|Publication number||US7010607 B1|
|Application number||US 09/661,214|
|Publication date||Mar 7, 2006|
|Filing date||Sep 14, 2000|
|Priority date||Sep 15, 1999|
|Publication number||09661214, 661214, US 7010607 B1, US 7010607B1, US-B1-7010607, US7010607 B1, US7010607B1|
|Inventors||William P. Bunton|
|Original Assignee||Hewlett-Packard Development Company, L.P.|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (11), Referenced by (434), Classifications (8), Legal Events (6)|
|External Links: USPTO, USPTO Assignment, Espacenet|
This application claims the earlier effective filing date of U.S. Provisional Application 60/154,150, filed Sep. 15, 1999, and entitled “Method and Apparatus for Transferring Data.”
1. Field of the Invention
2. Description of the Related Art
As information technology has matured, computing systems have evolved into what are now known as “enterprise computing systems.” An enterprise computing system is typically a large number of computing and storage devices, all of which are employed by users from a single concern, or “enterprise.” One popular type of enterprise computing system is an “intranet,” which is a computing system that operates like the Internet, but requires special authorization to access. Such access is typically only granted to employees and/or contractors of the enterprise. However, not all enterprise computing systems are intranets or operate along the principles of the Internet. One of the defining characteristics of the Internet is that communications among the computing devices utilize the Transmission Control Protocol/Internet Protocol (“TCP/IP”) as do intranets. However, there are many protocols, some of them proprietary, that may instead be employed in enterprise computing systems for, among other reasons, security purposes.
Thus, modern computer systems have advanced at a truly astonishing rate. Compute performance and memory capacity have followed Moore's Law predictions for over two decades. However, the predominant I/O sub-system architecture of these systems has not followed suite. Typical I/O subsystems have been designed around shared bus architectures that, while adequate in performance at the time of their introduction, have not been able to keep pace with the rapid advances in processor and memory subsystem performance. To meet the ever increasing I/O demands of the next decade, a revolutionary approach to I/O will be required that can co-exist with the evolutionary improvements to legacy I/O, such as Peripheral Component Interface (“PCI”) and Peripheral Component Interface Extension (“PCI-X”).
More particularly, Internet and electronic commerce has grown to the point where demands placed on existing computer systems are severely testing the limits of system capacities. Microprocessor and peripheral device performances have improved to keep pace with emerging business and educational needs. For example, internal clock frequencies of microprocessors have increased dramatically, from less than 100 MHz to more than 1 GHz over a span of less than ten years. Where this performance increase in inadequate, high performance systems have been designed with multiple processors and clustered architecture. It is now commonplace for data and software applications to be distributed across clustered servers and separate networks. The demands created by these growing networks and increasing speeds are straining the capabilities of existing Input/Output (“I/O”) architecture.
Peripheral Component Interconnect (“PCI”), released in 1992, is perhaps the most widely used I/O technology today. PCI is a shared bus-based I/O architecture and is commonly applied as a means of coupling a host computer bus (front side bus) to various peripheral devices in the system. Publications that describe the PCI bus include the PCI Specification, Rev. 2.2, and Power Management Specification 1.1, all published by the PCI Special Interest Group. The principles taught in these documents are well known to those of ordinary skill in the art and are hereby incorporated herein by reference.
At the time of its inception, the total raw bandwidth of 133 MBps (32 bit, 33 MHz) provided by PCI was more than sufficient to sustain the existing hardware. Today, in addition to microprocessor and peripheral advancements, other I/O architectures such as Gigabit Ethernet, Fibre Channel, and Ultra3 SCSI are outperforming the PCI bus. Front side buses, which connect computer microprocessors to memory, are approaching 1–2 GBps bandwidths. It is apparent that the conventional PCI bus architecture is not keeping pace with the improvements of the surrounding hardware. The PCI bus is quickly becoming the bottleneck in computer networks.
In an effort to meet the increasing needs for I/O interconnect performance, a special workgroup led by Compaq Computer Corporation developed PCI-X as an enhancement over PCI. The PCI-X protocol enables 64-bit, 133 MHz performance for a total raw bandwidth that exceeds 1 GBps. While this is indeed an improvement over the existing PCI standard, it is expected that the PCI-X bus architecture will only satisfy I/O performance demands for another two or three years.
In addition to the sheer bandwidth limitations of the PCI bus, the shared parallel bus architecture used in PCI creates other limitations which affect its performance. Since the PCI bus is shared, there is a constant battle for resources between processors, memory, and peripheral devices. Devices must gain control of the PCI bus before any data transfer to and from that device can occur. Furthermore, to maintain signal integrity on a shared bus, bus lengths and clock rates must be kept down. Both of these requirements are counter to the fact that microprocessor speeds are going up and more and more peripheral components are being added to today's computer systems and networks.
Today, system vendors are decreasing distances between processors, memory controllers and memory to allow for increasing clock speeds on front end buses. The resulting microprocessor-memory complex is becoming an island unto itself. At the same time, there is a trend to move the huge amounts of data used in today's business place to storage locations external to network computers and servers. This segregation between processors and data storage has necessitated a transition to external I/O solutions.
One effort to meet some of these demands resulted in what is known as the Future I/O (“FIO”) Architecture Specification. The FIO Architecture Specification lays the groundwork for this revolutionary approach to industry standard I/O components and their interconnections. It is an architectural successor to PCI and PCI-X that allows the I/O subsystem to scale proportionally with advances in the rest of the computer system. FIO supports both scaling in size to meet the needs of large enterprise systems designed to serve the Internet, and scaling over time to allow I/O performance upgrades to more closely track the up-grades of microprocessor and memory technology.
FIO is designed around a point-to-point, switched I/O fabric, whereby endnode devices (which can range from very inexpensive I/O devices like single chip SCSI or Ethernet adapters to very large host computers) are interconnected by one or several cascaded switch devices. The physical properties of the FIO interconnect support three predominant environments, with bandwidth, distance and cost optimizations appropriate for these environments:
The FIO switched fabric can be viewed as a reliable transport mechanism where messages are enqueued for delivery between end nodes. Message content and meaning is not specified by FIO, but rather is left to the designers of end node devices and software that is hosted on end node devices. Sufficiently reliable hardware transport protocols are defined to support both reliable messages (send/receive) and memory manipulation semantics (e.g., Remote Direct Memory Access (“RDMA”) and fetch operations (“fetchops”)) without software intervention in the data movement path. Adequate protection and error detection mechanisms are defined to allow messages to originate and terminate from either privileged kernel mode (to support legacy I/O and IPC needs) or user space (to support emerging IPC demands). The FIO Architecture Specification also addresses the need for a rich manageability infrastructure to support interoperability between multiple generations of FIO components from many vendors over time. This infrastructure provides inherent ease of use and consistent behavior in high-volume, cost-sensitive deployment environments. Also specified are industry-standard management interfaces that interoperate with enterprise class management tools for configuration, asset management, error reporting, performance metric collection and topology management necessary for data center deployment of FIO.
The FIO architecture has been carefully designed to minimize disruption of prevailing market paradigms and business practices. By simultaneously supporting chip, board and chassis interconnections, it is expected that vendors will be able to adopt FIO technology for use in future generations of existing products, within current business practices, to best support their customers needs. FIO is an I/O fabric connecting multiple media-specific I/O adapter devices. It is neither a competitor of nor replacement for any existing I/O media. FIO can support bandwidths that are anticipated to remain an order of magnitude greater than prevailing I/O media (SCSI, Fibre Channel, Ethernet). This ensures its role as the common interconnect for attaching I/O media. Reinforcing this point is FIO's native use of Ipv6 headers, which supports extremely efficient junctions between FIO fabrics and traditional internets and intranets.
FIO will be delivered by multiple vendors and channels supporting diverse customer needs and al-lowing customers the ultimate flexibility for purchasing computing components. The FIO Architecture supports implementations as simple as a single node computer system, and can be expanded to include: replication of components for increased system reliability, cascaded switched fabric components, and I/O adapters for additional I/O capacity and bandwidth, additional host node computing elements for scalable computing or any combinations thereof. FIO is a revolutionary architecture that enables computer systems to keep up with the ever increasing customer requirement for increased scalability, in-creased bandwidth, decreased central processing unit (“CPU”) utilization, high availability and support for Internet technology.
A second solution to this I/O problem has been proposed by the InfiniBand(SM) Trade Association. The InfiniBand(SM) Trade Association is an independent industry body that is developing a channel-based, switched-network-topology interconnect standard. This standard will de-couple the I/O subsystem from the microprocessor-memory complex by using I/O engines referred to as channels. These channels implement switched, point to point serial connections rather than the shared, load and store architecture used in parallel bus PCI connections.
The InfiniBand interconnect standard offers several advantages. First, it uses a differential pair of serial signal carriers, which drastically reduces conductor count. Second, it has a switched topology that permits many more nodes which can be placed farther apart than a parallel bus. Since more nodes can be added, the interconnect network becomes more scalable than the parallel bus network. Furthermore, as new devices are added, the links connecting devices will fully support additional bandwidth. This InfiniBand architecture will let network managers buy network systems in pieces, linking components together using long serial cables. As demands grow, the system can grow with those needs.
The trend towards using serial interconnections as a feasible solution to external I/O solutions is further evidenced by the emergence of the IEEE 1394 bus and Universal Serial Bus (USB) standards. USB ports, which allow users to add peripherals ranging from keyboards to biometrics units, have become a common feature in desktop and portable computer systems. USB is currently capable of up to 12 MBps bandwidths, while the IEEE 1394 bus is capable of up to 400 MBps speeds. A new version of the IEEE 1394 bus (IEEE 1394b) can support bandwidth in excess of 1 GBps.
However, one problem common to even these new approaches is how to “train” communications links within the computing system. “Training” refers to the process by which ports through which communications are conducted are brought into the fabric of the computing in a manner that is efficient, but not disruptive. The present invention is directed to resolving, or at least reducing, one or all of the problems mentioned above.
The invention comprises a technique for training links in a computing system. In one aspect, the technique includes configuring a first receiver in a first port using a first training sequence or a second training sequence; transmitting the second training sequence from the first port indicating the first receiver is configured; and receiving a second training sequence transmitted by a second port at the first port, the second training sequence transmitted by the second port indicating that a second receiver in the second port is configured. In a second aspect, the technique includes locking a communication link; handshaking across the locked link to indicate readiness for data transmission; transmitting information after handshaking across the locked link. And, in a third aspect, the technique includes transmitting a first training sequence from a first port and a second port; and synchronizing the receipt of the first training sequence at the first and second ports; transmitting a second training sequence from the first and second ports upon the synchronized receipt of the first training sequence at the first and second ports; and receiving the second training sequence transmitted by the first and second ports and the second and first ports, respectively, in synchrony.
The invention may be understood by reference to the following description taken in conjunction with the accompanying drawings, in which reference numerals identify like elements, and in which:
FIG. 50—Port States in Spanning Tree;
FIG. 51—Path Costs to Root (seen by Switch E);
FIG. 65—Simple Tree with Mixed Bandwidth Links and Adapter and Router Leaves;
While the invention is susceptible to various modifications and alternative forms, specific embodiments thereof have been shown by way of example in the drawings and are herein described in detail. It should be understood, however, that the description herein of specific embodiments is not intended to limit the invention to the particular forms disclosed, but on the contrary, the intention is to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the invention as defined by the appended claims.
Illustrative embodiments of the invention are described below. In the interest of clarity, not all features of an actual implementation are described in this specification. It will of course be appreciated that in the development of any such actual embodiment, numerous implementation-specific decisions must be made to achieve the developers' specific goals, such as compliance with system-related and business-related constraints, which will vary from one implementation to another. Moreover, it will be appreciated that such a development effort, even if complex and time-consuming, would be a routine undertaking for those of ordinary skill in the art having the benefit of this disclosure.
The present invention is a “link training” technique for use in SANs implementing, e.g., Future I/O or InfiniBand I/O technologies. The following detailed description is separated into three parts. First, there is a discussion of one particular embodiment of a SAN implementing an FIO system in accordance with the present inventionSecond, there is a more general discussion of various aspects of the invention that are illustrated in
FIO defines a switched communications fabric allowing many devices to concurrently transfer with high bandwidth and low latency in a secure, remotely managed environment. Endnodes can communicate over multiple FIO ports and utilize multiple paths through the FIO fabric. The multiplicity of FIO ports and paths through the network are exploited for both fault tolerance and increased bandwidth for data transfers. FIO hardware off-loads much of the CPU and I/O adapter communications overhead. This allows multiple concurrent communications over a switched network without the traditional overhead associated with communicating protocols. The FIO SAN gives its I/O and IPC clients zero processor-copy data transfers, with no kernel involvement, and uses hardware to provide completely reliable, fault tolerant communications.
Turning now to
Each of the host processors 110 includes a plurality of CPUs 112, at least one system area network interface controller (“SANIC”) 114, and some memory 116. The CPUs 112 may be any suitable processor known to the art, such as an Intel x86 processor, e.g., a Pentium III or Pentium Pro processor. The SANICs 114 interface between the FIO fabric and the host processors 110. The memory 116 may be private, shared, or distributed depending on the particular implementation.
The RAID subsystem 120 also includes a processor 112, a memory 116, and a SANIC 114. Enterprise computing networks, however, typically use mass storage units extensively, and RAID is common. As is known in the art, this technology uses multiple disks in a single unit in conjunction with one another to improve performance and enhance reliability. A RAID improves both performance and reliability by storing data onto multiple physical disk drives. Reliability is improved through redundancy of information, and performance is improved by exploiting the fact that multiple disks are available for use.
Note that there are several alternative, standard RAID implementations as well as many derivative or proprietary versions. The standard alternatives include:
Each of the FIO I/O Adapter backplanes 130 includes an embedded FIO switch 140 interfacing with a plurality of FIO adapter cards 132. The adapter cards 132 may interface with devices operating in accordance with practically any protocol, including the Small Computer System Interface (“SCSI”), Fibre Channel, Ethernet, protocols or some sort of graphics application. The FIO I/O adapters can range in complexity from single Application Specific Integrated Circuit (“ASIC”) FIO attached SCSI adapters to large memory rich RAID boxes that rival a host in complexity.
More generally, the FIO system 100 is made up of host and I/O adapter “endnodes” communicating through FIO “links.” Endnodes include the host processors 110 or the I/O adapters 130 that connect to the FIO fabric 115. An endnode has Queue Pairs (discussed further below) that are the source or destination of FIO frames (also discussed below). Endnodes generate request frames and return acknowledgment frames. Links attach endnodes to other endnodes or, more typically, attach endnodes, switches, and routing devices. Links can be copper or optical cables or printed circuit wiring on a backplane or board. Links connect endnodes to switches and switches to switches and are defined for copper trace and cables. FIO also defines a board form factor for FIO connected I/O adapters. Since FIO is intended as a general purpose I/O standard to supplement PCI, FIO architecturally defines an I/O card suitable for plugging into a backplane. If vertically mounted, the card can fit in a 3U cage. Each of these cards has one require FIO backplane connection with an optional second FIO port. A dual height 6U version (with 2 to 4 FIO ports) is also defined. See the FIO Electrical Mechanical specification for a full description.
More precisely, an endnode (either a host processor or I/O adapter) can have one or more SAN Integrated Circuits (“SANICs”). The SANIC can in turn have one or more connections, or SAN Link Level Engines (“LLEs”), attached to the FIO fabric.
There are several different FIO defined terms for naming and identifying an endnode:
The Queue tail, consists of a Send Queue and a Receive Queue, message requests are initiated by posting Work Queue (WQE) to the Send Queue. The FIO client's message is referenced by a gather-list in the Send WQE as shown in
Hardware reads the WQE and packetizes the messages into frames and flits. Frames are routed through the network, and for reliable transport services, are acknowledged by the final destination. If not successfully acknowledged, the frame is retransmitted by the source endnode. Frames are generated and consumed by the source and destination endnodes, not by the switches and routers along the way.
Flits are the smallest unit of flow control on the network and are generated and consumed at each end of a physical link. Flits are acknowledged at the receiving end of each link and are retransmitted if there is an error.
The send request message is sent as two frames. Each request frame is in turn broken down into 4 flits. The ladder diagrams shown on
The FIO SAN is a fully switched fabric. There is no concept of an FIO bus or an FIO arbitrated loop. The FIO fabric is made of subnets connected to each other with FIO routers. A subnet has up to 216 endnodes, switches and routers. A subnet is a group of endnodes and cascaded switches that is managed as a single unit. Typically a subnet would occupy a single geographic or functional area, for example, a single computer system in one room. The switches in a subnet can do very fast worm-hole or cut-through routing of FIO messages. The switch examines a 16-bit destination local ID (DLID) that is unique within the subnet. This allows the switch to quickly and efficiently route incoming FIO message frames. The switch is relatively simple, typically implemented as a single ASIC. By cascading switches, a subnet can have hundreds or thousands of endnodes. For expansion to much larger systems, subnets are tied together with FIO Routers. The FIO router is able to interpret the Ipv6 destination ID and therefore route the IP-like FIO frame.
Switches and Routers both degrade gracefully when links are over-utilized. Link level backpressure is used to temporarily slow the flow of data when multiple input frames compete for a common output. Link or buffer contention does not cause the loss of data.
Switches, routers and endnodes use the FIO link protocol to transfer data. The link protocol supports an automatic error retry. Link level acknowledgments detect errors and force retransmission of any data impacted by bit errors. Link-level error recovery greatly reduces the number of data errors that are handled by the end-to-end protocols. Note that whether the error is detected and corrected by the link-level or the end-to-end protocol, the FIO user client is not involved with error recovery.
Switches and Routers use multiple virtual lanes within a single physical FIO link. Frames injected into the FIO fabric follow a particular virtual lane from their source to their destination. At any one time, only one virtual lane can make progress on a given physical link. Virtual lanes are a technique where link level flow control can be applied to one lane without affecting the others. When a frame on one virtual lane blocks (due to contention or QOS considerations) a frame on a different virtual lane is allowed to make progress. Virtual lanes are used for several reasons:
Thus, there may be a variety of paths through the FIO subnet. Within a subnet, a path from a source port to a destination port is determined by the LID (local identifier) of the destination SANIC port. Between subnets, a path is determined by the Ipv6 address of the destination SANIC port. The paths used by the request frame and its corresponding acknowledged (“Ack”) or not acknowledge (“Nack”) frame are not required to be symmetric. Switches select an output port based on the DLID (i.e., oblivious routing). A switch may use one set of routing decision criteria (e.g., one routing table) for all its input ports or it may use a separate set of criteria for each input port. Each port on an endnode may have multiple Ipv6 addresses. Multiple addresses may be used for several reasons:
A Host FIO endnode and an I/O adapter endnode will typically have different capabilities. For example, a host endnode might support 4 FIO ports, hundreds of thousands of QPs, and allow incoming RDMA operations. An FIO attached I/O adapter endnode might only support one or two FIO ports and tens of QPs and not allow incoming RDMA operations. Furthermore, a low end FIO attached I/O adapter may choose to use software to handle much of the network and transport layer functionality as the Host FIO endnode does in hardware. FIO allows significant freedom of implementation while still defining a common architecture to ensure interoperability. The FIO architecture does not define many aspects of the endnode or switch design. This is done to allow flexibility to meet specific design goals (e.g., cost and performance) and to encourage competition and new designs. To start the FIO design efforts, the FIO Coreware group defines several FIO implementations, a pre-silicon simulation environment and validation suites. The coreware defines internal hardware inter-faces that allows portions of the coreware design to be refined and reused.
Channel and Memory Semantics. FIO supported communications give the user both channel semantics, as shown in
FIO Scope. The FIO Architecture Specification defines the FIO interconnect fabric, routing elements, endnodes, and the formats and protocols for communication. It does not specify specific I/O commands or cluster services. These will be implementation specific depending on factors that will be apparent to those skilled in the art having the benefit of this disclosure. For example, consider an FIO attached SCSI adapter. FIO does not define the disk I/O commands, how the disk communicates with the SCSI adapter, or how the operating system (“OS”) views the disk device, or which processor in the cluster owns the disk adapter. FIO is an essential underpinning of each of these operations, but does not directly define any of them. Instead, FIO defines how data and commands are transported between the host processor and the SCSI adapter. FIO handles the data communications for I/O and IPC in a multi computer environment. It supports the high bandwidth and scalability required for I/O. It also supports the extremely low latency and low CPU overhead required for IPC. With FIO, user clients can bypass the operating system (“OS”) kernel and directly access FIO network communication hardware, enabling efficient message passing protocols. FIO is well suited to current computing models and will be a building block for new forms of I/O and cluster communication. FIO allows I/O adapters to communicate among themselves or communicate with any or all of the processor nodes in a system. With an I/O adapter attached to a SAN, the I/O adapter has the same communications capability as any processor element.
Queue Pairs. The FIO client's software communicates with FIO hardware through one or more “queue pairs”. A SANIC in a host processor will support many thousands of QPs while the SAN interface in an I/O adapter might support less than ten QPs. As is shown in
Specifically, the WQEs supported for Send Queues are:
FIO operations use virtual addresses and existing virtual memory protection mechanisms to assure correct and proper access to all memory. FIO application are not required to use physical addressing for any operations. Remote memory accesses are further restricted by the user mode client of FIO. Before a process allows remote memory accesses, it first binds a remote memory access key to a description of the buffer. Only an incoming RDMA request with the correct key can gain access to the specific range of previously defined data buffers. This binding of memory access permissions can be done entirely in hardware and does not require any CPU overhead or kernel involvement. Remote accesses can be controlled with byte granularity on both the lower and upper bound of each virtually contiguous buffer.
FIO Transport Services. The FIO transport mechanisms provide four types of communication services. When a QP is created, it is set to provide one of the four types of transport services:
A connected service requires a process to create a QP for each process it wishes to communicate with over FIO. If M processes on N processors all wish to talk, each processor requires M*N2 QPs. A process can connect a QP to a QP on the same SANIC as itself.
The Reliable Connection is reliable because hardware can maintain sequence numbers and acknowledges all frame transfers. A combination of hardware and SAN driver software retries any failed communications. The client of the QP sees reliable communications even in the presence of bit errors, receive buffer under runs, network congestion, and if alternate paths exist in the fabric, failures of fabric switches or links.
The FIO acknowledgments are used to ensure data is delivered reliably to the remote QP or memory space. The acknowledgment is not a process level acknowledgment—it doesn't validate the receiving process has consumed the data. The acknowledgment only means the data has reached the destination. The Reliable Datagram service is connectionless. It allows the client of the QP to communicate with any other QP on any other remote node. On the Receive Queue, it allows in-coming messages from any QP on any other remote node.
The Reliable Datagram service greatly improves the scalability of FIO. Since the Reliable Datagram service is connectionless, an endnode with a fixed number of QPs can communicate with far more processes and processors compared to that possible using the Reliable Connection transport service.
The Reliable Datagram service uses sequence numbers and acknowledgments associated with each message frame to ensure the same degree of reliability as with the Reliable Connection service. The SANICs maintain end-to-end specific state to keep track of the sequence numbers, acknowledgments and time-out values. The end-to-end state is shared by all the connectionless QPs communicating between a pair of processors.
The Unreliable Datagram service is also connectionless. It is used by management applications to discover and integrate new FIO switches and endnodes. It does not provide the reliability guarantees of the other services, but also can operate with less state maintained at each endnode.
The fourth service is not technically a transport. It allows the QP to send and receive raw datagram frames. With the raw datagram service, the FIO hardware can generate and consume standard protocols layered atop Ipv6, such as the Transmission Control Protocol (“TCP”) and the User Datagram Protocol (“UDP”). The FIO frame header can be mapped directly to and from an Ipv6 header. Native Ipv6 frames can be bridged into the FIO SAN and delivered directly to a QP. This allows an FIO client to support any transport protocol running on top of Ipv6.
The FIO client can register with the hardware in order to direct datagrams for a particular upper level protocol (e.g., TCP) to a particular QP. FIO hardware can de-multiplex an incoming Ipv6 stream of datagrams based on the Next Header field as well as the destination IP address.
FIO Transactions. An FIO data transaction is composed of several hardware and software steps. A user-mode or kernel-mode process can be a client of an FIO data-transport service. The process accesses the FIO hardware through one or more QPs (queue pairs). The process calls an operating-system specific programming interface called the FIO Verbs. The code implementing the FIO Verbs in turn posts a Work Queue Element (WQE) to the QP work queue.
The method of posting a WQE and the WQE format are not defined by FIO, and are implementation specific. This allows innovation in implementations as well as designs that support various cost performance points. This lack of standardization in no way affects interoperability. A user process talks to the FIO Verbs in a well-defined manner and the format and protocols of the data going across the wire is completely specified. FIO allows devices to interoperate in a heterogeneous vendor environment.
The hardware detects the WQE posting and accesses the WQE. The hardware translates and validates the WQE's virtual addresses and accesses the data. The outgoing message buffer is split into one or more frames. To each frame the hardware adds a transport header (sequence numbers, etc.) and a network header (destination IP address or destination Local ID, etc.). The appropriate local or global network header is added depending on if the destination resides on the local subnet or on a remote subnet.
A frame is the unit of information that is routed through the FIO fabric. The frame is an endnode-to-end-node construct, in that it is created and consumed by endnodes. Switches and routers do not generate or consume either request or acknowledgment frames, they simply move them closer to the ultimate destination (although routers do modify the frame's network header when the frame crosses a subnet boundary). When traversing a subnet a single frame stays on a single virtual lane.
When the frame is placed onto the wire, it is further broken down into flits. The flit is a unit of link-level flow control. The flow of flits is subject to the link-level protocol which can do flow control or retransmission after an error. The flit is a link-level construct. It is created at each endnode, switch or router output port and consumed at each input port. The header and payload sizes for FIO transactions are set forth in Table 1.
Header and Payload Sizes
(exclusive of all
variable, in byte
increments up to
4 KB max
16 Byte (Send
Request) 24 Byte
(RDMA Request) 8
variable, in 16 B
increments up to 512
If a reliable transport service is in use, after the frame reaches the destination, the destination sends an acknowledgment frame back to the sender. The acknowledgment frame is key to the reliable transfer of data. It allows the requestor to validate that the request reached the destination. An acknowledgment is sent back to the requestor after each request frame. The requester may have multiple outstanding requests before it receives any acknowledgments. This window size is agreed to when the QP is created.
Structure of the Layered Architecture
The FIO architecture is described as a series of independent layers. Each layer is independent of the other layer and has a well defined interface to the layer above and below. Independent layers are an effective way to define an architecture. They allow modifications to one part of the design without affecting the other.
The Physical Layer. The Physical layer provides definition those element that allow for interoperability between components from ASIC OEMs, I/O adapter IHVs, and system vendors. The elements addressed are Electrical Signaling, Mechanical Form Factors, Power Control/Hot Swap and Hardware Management. The FIO Electrical/Mechanical Specification describes the Physical Layer.
Links are point-to-point in structure and are used to connect endpoints to switches and switches to switches. A link is a full-duplex interface using small swing differential signaling and is comprised of 40 conductors (10 data bits*2 differential*2 directions). The data on the link is encoded using an 8b/10b algorithm; this guarantees a maximum time between transitions on the lines and thus allows for clock ex-traction from the data stream using one or more phased locked loops (PLL). The link interface is definitionally parallel and is self tuning for jitter and both inter- and intra-pair skew compensation. The specification calls for a roadmap of interoperable link generations. The first uses signaling rate of 1.25 Gb/s per bit yields a net payload bandwidth of 1 GB/s per direction (2 GB/s of cross-sectional band width). The second generation calls for 2.5 Gb/s per bit for a net payload of 2 GB/s per direction. The third generation calls for 5 Gb/s per bit for a net payload of 4 GB/s per direction. As the 5 Gb/s signaling rate is recognized to be pushing the edges of foreseen technologies, pins are reserved to allow for growth to a 48 conductor (12 data bits*2 differential*2 directions) interface.
The targeted electrical topologies include on-board (e.g., switch to switch on a board), slot on board (e.g., switch to adapter like PCI), board to board (e.g., pluggable switch to pluggable adapter), remote copper attachment (e.g., modular I/O expansion), and remote optical attachment (e.g., datacenter expansion). This topologies bring a combination of FR4 PCBs, board connectors, assembled cable and connectors, and optical components. The differential signaling used enables the ability to drive copper with distance with still being applicable on PCBs. The encoding provides for DC balance in the data stream such to support optical implementations straightforwardly.
Standard slots with connector definition and board form factors are specified. The board sizes are based on IEEE and EIA industry standards. Specifically, three boards are defined: “Single” which measures 100 mm×220 mm and support either 1 or 2 link ports; “Double” which measures 233.35 mm×220 mm and supports 1, 2 or 4 link ports; and “Half” which measures 47 mm×220 mm and supports only 1 link port. Half sized board are intended for cable attachment to backplanes and optical conversion functions. Thermal and ElectroMagnetic Interference (EMI) properties are also addressed. Chassis mechanicals, other than those required by formation of the standard slot, are not defined.
Future I/O defines that adapter boards have power conversion circuits to take a 12 V distributed primary supply and produce those voltage levels required by an implementation. This allows for adapter ASIC technology evolution without implications on this specification. In conjunction with these power conversion circuits, pin sequencing and timing requirements facilitate both hot insertion and removal. Additionally, a separate auxiliary, low power input is defined to allow for wake-on-media and other hardware management operations when the primary inputs are not enabled.
Hardware Management defines those mechanisms that allow for control and monitoring of physical functions from within the Future I/O network and from the system chassis. Physical functions includes elements like power, thermal, vital product data (VPD), The “in-band” mechanisms defines access to This enables a broad range of management capabilities across differing system designs.
The Link Layer.
The details of the various layers up to Transport Layer and Verbs are described in various chapters of this specification. The Upper Layer Protocols are the “application level” processes which employ the Future I/O Transport services for data movement, and include, for example device drivers and their corresponding I/O adapters, or user-level distributed message-passing applications.
This chapter describes the functions handled by the Link Layer of the Future I/O Architecture. The Link Layer includes 2 functionally separate sub-layers: the Link Flit Level Logic, and the Link Encoding Level Logic. The Link Flit Level Logic handles sending and receiving of data across the links at the flit level. The Link Flit Level Logic is independent of the Physical Layer, other than the dependency on link transmission bandwidth (1–4 GB/sec). The Link Encoding Level Logic provides the digital data encoding/de-coding and multiplexing/demultiplexing necessary to interface the Flit level logic with the particular Physical Layer implemented.
This document describes only those portions of the Physical layer which are implemented within Future I/O Component chips (e.g., SANIC, Switch, or I/O adapter ASICs). These portions primarily include all the digital logic which is dependent on the particular Physical Layer implementation used. The other portions of the Physical Layer, which include the analog functions within a chip (e.g., drivers and receivers, de-skew logic, etc), as well as the interfaces between chip (e.g., voltage, attenuation, and power levels, connectors, board trace rules, cable specifications, etc.) are all incorporated into the Future I/O Electrical/Mechanical Specification, which can be found separately.
The Flit level protocol handles operations dealing with link-level flow control, link-level error control, and distinguishing between different types of data transferred over each link. The Link Flit layer logic relates only to the link-level protocol, i.e., each link-level engine is independent of all other link-level engines in the system. The Link Flit level protocol handles several functions. Link error control monitors the transmitted data for transmission errors, and corrects errors which can be fixed by data retransmission. Link data separation allows the link protocol logic to distinguish different types of data sent over the link, and provides some independence between the different types of data and control bytes flowing across the links. Finally, the link flit level handles link-level flow control, assuring that data is not transmitted across each link until there is guaranteed space to receive it into at the link receiver. The Link Flit level logic incorporates the following primary features. * Frames, which are the unit of routing and end-to-end recovery, are split at the links into 1 or more flits. Flits (flow control units) are the fundamental unit of flow control, error control, and transmission granularity. This mechanism allows a combination of fine granularity for switching and link-level protocol, with larger frame-level granularity for protocol processing and the source and destination endnodes, for protocol-processing efficiency. The support for VLs on all links also provides fundamental link-level support for a wide variety of quality of. service or transmission priority capabilities, and allows the link level logic to be nearly independent of frame length, with minimal interaction between large and small frames on different VLs.
All transmission over each link of the Future I/O network is executed in units of a “Flit”, which is the in-divisible unit of flow control and error control in the network. (The term flit is derived from “flow control unit”.) All flow control through the network is executed in units of flits, and any link-level error detection and error recovery procedures are executed in units of flits.
The flit is of a variable size and has its own flit description, error control, and flow control information. One or more flits make up a frame, which can contain from 0 B to many KB of frame payload. The routing information for the data is kept at the frame level, and not at the flit level. All the flits for each frame are routed through the network along the same path, and are efficiently reassembled into a single per-VL frame buffer at the receiving endnode.
Flits of different frames can be interspersed on each link, at flit boundaries, provided that the frames are on different VLs. Within a particular VL, sequential flits must always be part of the same frame, on each link of the network. These rules assure that at the receiving endnodes, frames appear as fully reassembled frames in each of a set of per-VL receive buffers. This link-level segmentation and reassembly function simplifies and optimizes protocol processing for and interface to the higher-level endnode protocols.
The primary reason for using flit-based transmission on Future I/O links is to provide efficient transmission for both very small (8B) and very large (multi-KB) frames across a large switched fabric, with support for multiple priority levels. Large frames are a great help for the endnode engines for obtaining high-band-width data transmission, to minimize the per-frame processing overhead at the endnodes, particularly the receiving endnode. However, the link-level transmission and switching protocol and implementation be-comes more efficient with smaller transmission units, to minimize link-level buffer sizes, lower switching latency, and allow efficient interspersing of high- and low-priority traffic. Splitting long frames into shorter flits allows the endnodes to operate on a frame-level granularity, for better endnode protocol processing efficiency, while links and switches operate on a flit-level granularity, for better link buffering efficiency, lower latency, and faster switching between priority levels.
The flits are separated by 8B flit delimiters, which contain information about the subsequent flit, and about the error control and flow control state of the link. Each flit header also contains an error detection CRC which covers the flit header fields as well as the body of the *previous* flit.
Conceptually, the flits and flit delimiters flow across the links as shown in
Each flit includes an 8B header, and between 0 and 512B of flit body, in multiples of 8B. This relatively large granularity simplifies data flow within the high-speed interfaces in the switches and endnodes. The 512B maximum flit size assures that the maximum time between arbitration across VLs is limited to 0.5 microseconds or less. Larger flits (e.g., 1 KB or larger) would allow a negligible amount of extra transmission efficiency. Frame payload may be defined to 1 B granularity, as indicated in the frame's local routing header, with filler bytes inserted to round out the final flit of the frame to the nearest larger 8B flit boundary.
Frame data is tightly packed into flits, starting a frame with a flit start, filling each until the end of frame, at which point a partly filled flit (and chunk) may occur. The Frame is assumed to carry its own length, so the link layer need not identify the length of a partly filled flit.
The end-point generating the frames determine the size of the flits used to transmit the frames. There is no segmentation or reassembly within a flit by the switches, in order that the flits generated at the source end-point are the same size as the flits received at the destination end-point.
When used to schedule transmissions, a flit is handled in its entirety. In general, flit scheduling between different VLs is done on a per-flit basis, regardless of the size of the flits transmitted, so endpoints that send small flits will tend to get less than their expected bandwidth in bytes per second, even though their percentage of flits is correct.
Components must allocate buffers and manage credits as though the flits are of a particular size, 64 bytes of data. If an actual flit is smaller, then it still uses a buffer and consumes a single credit. If a flit is larger than 64 bytes (data) it must consume more than one buffer and credit. Each buffer must carry the extra information associated with a flit; its data CRC, type, VL etc. (Note that some of this may be implied by the buffer itself, such as the VL. The opcode may also be reduced to a single bit such as start of frame.
The buffer and credit management size of 64 bytes was chosen to allow a better utilization of endpoint buffers and credits in the face of short flits. If we always manage the buffers as full flits, then 3000 bytes of buffer space contains about 6 flits. If we send very short flits on the wire, we may quickly starve the system for credits when the link is more than a few meters long.
To recover from soft errors, such as bit transmission errors on noisy links, flits can be re-transmitted at the link level between ends of a single link. Each flit containing valid information is assigned a sequence number, and is buffered at the link sender, for retransmission in case it is not acknowledged.
The originator of a flit keeps track of which flits have been acknowledged by looking for incoming ‘Link Ack/Idle’ flits. The sequence number in the Link Ack/Idle flit specifies the next expected flit, rather than the sequence number of the flit being acknowledged. This allows an ‘aggregation’ effect such that a single Link Acknowledge/Idle flit can be acking more than one flit.
The protocol allows for re-transmission of flits that may have already been received (i.e., a Link Ack/Idle flit was lost). Rather than attempting to get only the bad flit re-transmitted, the protocol requires that everything be sent in sequence; the unacknowledged flit and everything sent after it is sent again in the same order as before. This simplifies the link receiver logic considerably, since it only has to determine whether the incoming flit's sequence number is the expected number or not.
The receiver is required to discard any valid flit whose sequence number does not match the current expected number, signal the Link Encoding Layer to generate and generate a Link Ack/Idle flit in response, regardless of whether the incoming flits sequence number is less than or greater than the expected sequence number.
Flit types are as shown in
Virtual Lanes provide independent, non-deadlocking data flows through the network. Each virtual lane has its own buffering and flow control, and the tunable VL arbitration mechanism allows the VLs to carry different priorities of traffic, as well. There are sixteen virtual lanes architected in the flit headers, but FIO components may be implemented with fewer VLs supported. VL 15 is reserved for management/network control functionality, for transmit-ting connection management, resource management, and subnet fabric management control frames, and shall be supported by all components. All components are expected to support at least VL0–VL7, as well as VL 15. Frames transmitted between end-points may be transmitted on any VL supported by all components on the path between those end-points. The end-points must agree on which VL to use, based on traffic priority and deadlock avoidance considerations. Switches and routers shall not change VLs during transmission, to prevent potential deadlock conditions.
The SEQ field allows individual flits to be distinguished, so that link-level error recovery can be used to correct link errors on a flit-by-flit basis. All flits except Idles and link initialization flits are sequenced, with sequence numbers in the 16-bit range from 0x0000 to 0xFFFE. The “all 1s” number, 0xFFFF, is reserved for distinguishing flits that are unsequenced—flits with SEQ=0xFFFF are not buffered at the link sender, and can not be retransmitted.
The link receiver logic tracks an expected sequence number, and only accepts the flit with the expected sequence number for transmission to the higher-level logic. If a flit with a sequence number other than the one expected is received, it is discarded silently.
The size requirement for the sequence number field is determined by the sizes of the transmitted flits and by the link BW*Roundtrip time supported. With an 16-bit sequence number field, up to 65535 flits can be outstanding. Assuming a minimum sequenced flit size of 24 bytes (the size of an End-to-End Ack Frame), this field covers at least ((2**16)−1)*24B=1.57 MB. Since each meter at 1 GB/s and 5 ns/meter supports 10B, this size of sequence number supports up to 150 km of link distance. Supporting this link distance* bandwidth will require single-mode optical fiber technology, which is not described in the current version of the Future I/O electromechanical specification, but which is expected to be at some time in the future.
This field describes the size of the flit body following the flit header, measured in 8-byte chunks. The actual flit size is 8 bytes times the value in this field, with the exception that the value 0b000000 may be interpreted as 64, for some TYPE values, to allow flits with up to 512 bytes (8*64 bytes) in size.
The link level credits are managed in units of 64 bytes of flit body only (a flit with a body shorter than 64 bytes is expected to use a full buffer, a longer flit uses multiple 64-byte buffers). The credits are managed as relative credits, the credit field in the flit is added to the available credits at the transmitter at each flit. Each valid normal flit contains a credit field, which may be set to zero.
The credits are forced to zero for each VL at link initialization, and credit-only flits are sent following link initialization to notify how much receive buffer space is available for each VL.
Credits are only transmitted in sequenced frames, which are buffered at the sender, acknowledged by the receiver, and retransmitted if they are not acknowledged. This assures that link credit is protected from loss to the same extent that the rest of the flit delimiter field data is protected. This allows incremental rather than absolute link credit accounting, which allows more flexibility in link technology design for bandwidth and distance. For current multi-mode parallel fiber modules, the maximum link bandwidth*roundtrip time is in the range of 3,000B, which would be covered by a credit counter with a size of (ceiling(log 2 (3,000/64))=) 6 bits.
The link level protocol relies on three independent, efficient error detection mechanisms for detecting link transmission errors, in addition to the implicit error detection mechanisms (e.g., valid sequence number, valid TYPE, valid VL, etc.). These error detection mechanisms include the following.
Note that the LCRC is independently generated on each flit, since it covers link-level delimiter information (link credit, link flow control, etc.), as well as flit body (Frame header and Frame Data). If the link receiver gets an LCRC check, all of the other fields in the flit delimiter must be considered suspect and ignored, so no further error checking is possible.
Note also, that the ECRC calculation starts at the beginning of a frame, and is calculated cumulatively across the frame. That is, a running 16-bit ECRC is calculated along the frame bytes, and “snapshots” of this running ECRC calculation are inserted into the delimiter of each flit, as it is sent out. ECRC calculation is re-started from 0 at the beginning of the following frame. There is no necessity for a separate frame-level header CRC or data CRC.
ECRC is cumulatively collected separately for frames on each VL, requiring separate accumulators for each VL to allow the traffic from different lanes to be intermixed. The ECRC accumulator is restarted at frame boundaries.
If there is no data portion present in the flit before the delimiter, the ECRC is set to Zero, and the LCRC is calculated on the ECRC field and the other 4 delimiter bytes only.
For any particular flit slot across the link, the sending link layer may have a choice of flits to send.
The first phase of priority arbitration determines whether Time-of-Day Service Flits are to be transmitted. These flits, which allow a globally-accurate time-of-day to be maintained across all nodes of the network, must be sent immediately as they are generated, for global time accuracy.
The second phase of priority arbitration determines whether a link needs to be acknowledged, to assure the opposite end of the link that flits have been received correctly (in case of return ack time-out), or if the link needs to be re-initialized, by sending init flits. This level assumes that the more basic level of link initialization (tuning, de-skew, and length measuring using Ping and Pong flits) have been accomplished.
The third phase of priority arbitration determines whether any control frames originated from this switch or SANIC need to be sent. If the switch output port is not currently transmitting a packet on any VL, a locally generated control frame will have priority over all new packets. Once a service packet has left the chip that created it, the service packet will be handled like any other packet.
The fourth phase of priority arbitration determines which of the packets on the VLs of the higher level interface will be sent. This VL arbitration mechanism is described further below.
The fifth phase of priority arbitration determines whether there are credits to send and whether a new RSEQ needs to be sent. If credits need to be sent a credit-only flit is sent.
Finally, if no credits or other flits need to be sent, an Ack/Idle flit is sent.
There is an arbitration between the VLs at each link sender for transmission priority, since, at any time, there may be data on any or all of the VLs ready to be sent. This includes link senders in SPEs in SANICs, adapters, or switches. A wide variety of arbitration mechanisms are possible (round-robin between VLs, absolute priority per VL, guaranteed bandwidth per VL, etc., or various combinations), and the reality is that no arbitration mechanism is optimal, for all possible traffic patterns. The optimal VL arbitration mechanism at each link sender is determined by the overall quality-of-service level objectives provided to the various end-to-end data streams, which is a network-wide configuration, and depends strongly on the type of data being sent by the various end-points.
To handle this situation, the VL arbitration engines in each link sender will have to be configurable, to allow the fabric management to optimize the arbitration between VLs to match the overall quality of service objectives for the network. We are evaluating various VL arbitration configuration mechanisms, which will be configurable based on values programmed into configuration registers in each link sender.
For the entire network to work efficiently, the VL arbitration engines in each link sender need to be coordinated together under an overall fabric QoS policy, which is described further below.
The following counters are reset to zero at power-up, and count until reset by software. All values read are 64-bits so it is possible for software that periodically checks these counters to avoid resetting them by taking the difference between each reading as the current number. The minimum set of link level statistics counters is listed below.
Each sender maintains a single Link Retry timer (per FIO port) to control the interval for re-sending flits, or taking down the link if there's no response.
Future I/O employs a mechanism for synchronizing the FIO link between the devices on either end of the FIO link as follows:
A requirement for Future I/O is to effectively support future improvements in link transmission bandwidth. As multiple generations of Future I/O link level technology are designed and built, a common procedure for allowing interoperation of Future I/O components of different generations together. In more detail, as senders and receivers supporting a subset of the possible link speeds of 1 GB/s, 2 GB/s, and 4 GB/s, each sender/receiver pair must negotiate together to determine the highest commonly-supported link speed, and negotiate inter-operating at that speed.
This negotiation could be carried out locally, i.e., the sender/receiver negotiation to highest common link speed could be carried out in hardware, without knowledge of higher level layers. Alternatively, the link protocol hardware could be simplified by only requiring the link protocol hardware to determine the highest commonly-supported link speed, while allowing the fabric control service decide which speed each link should operate at.
In Future I/O, the plan is to allow the fabric control services to decide which links operate at which speeds, up to their highest commonly-supported speeds. In addition to link protocol design simplicity, this choice offers the higher-level logic to make global routing decisions, in order to assure load-balancing across the network links. Doing effectively load-balancing requires network-wide knowledge, so the service control logic to assure that most of the transmission bandwidth is distributed across faster links.
The general scheme for operating a link at higher than the base speed is as follows:
The Future I/O Link Encoding and Link Physical Logic are the lowest layers of the of the FIO protocol stack. These layers provide both the logical and physical conversions necessary to mask the attributes of the physical media for the Link Flit Layer protocol. This masking allows the Link Flit Layer protocol to communicate peer-to-peer, using flits, with each flit made up of one (or more) eight-byte-words (EBW) of unencoded data. The first generation (GI) definition of the FIO link encoding and link physical logic layer transfer data at a rate of 1250 Mbyte/sec, across a copper or optical link composed of 10 parallel lines, each running at 1.25 Gbit/sec. Future generations of FIO will transfer data at two and four times the first generation possibly using other physical link technologies. Second and third generation physical links may require changes to either or both the link encoding layer or the link physical logic layer. These changes will be transparent to the Link Flit Logic and other upper layers of the FIO protocol stack. All further discussion in this chapter refers to this first-generation physical link.
The primary functions provided by the Link Encoding Layer are to:
3) Conversion between digital and analog signals, The Physical Logic portion of the Link layer is described in the Future I/O Electro-Mechanical Specification, which will be available separately as described in the preface to this document.
The transmitter logic portion of the link encoding logic transforms the 1250 Mea-Bytes/second eight byte wide flit data stream to a ten byte wide stream; each of the ten streams are encoded 8b-to-10b and serialized and transfer at 1,250 Mea-bits/second to the receiving chip. Logic on the receiving chip de-serializes and de-codes the ten bit streams, which are converted to the eight byte wide flit data stream for next higher layer.
The FIO link encoding layer interface the upper layers of the Link Layer to the FIO physical media. At the higher layers data moves as flits, each flit composed of one or more 8-byte words. The FIO physical media is made up of 10 parallel serial links. The link encoding layer implements a multi stage conversion to sup-port peer-to-peer communication at the link flit interface. The data format conversions are listed in transmit order below:
The FIO link encoding uses the industry standard 8b/10b code invented by IBM (U.S. Pat. No. 4,486,739, Peter A. Franaszek and Albert X. Widmer. “Byte Oriented DC Balanced (0,4) 8B/10B Partitioned Block Transmission Code” (Dec. 4, 1984).), and used by Fibre Channel, Gigabit Ethernet (IEEE 802.3z), FICON, and ServerNet. The 8b/10b code was selected to provide DC balance, limited run lengths, byte (symbol) and word synchronization capabilities, and ability to distinguish between data characters and control characters. The code conversion for both data and control characters is shown in
Link commands are uniquely identifiable code groups or ordered sets used for low level link control. Link commands are inserted from the link data stream before the transmitter's encoding logic and removed-following the receiver's decoding logic. Link layer commands are generated by or presented to the flit layer of the protocol; commands which do not force training or synchronization are transparent to the flit byte stream. Link commands are inserted simultaneously on all 10 byte streams creating 10 parallel copies of the same of the same command. Some portions of this section are closely tied to subject matter in the Future I/O Electromechanical specification, and refer to facilities described in more detail there. Please refer to that specification for more detailed information.
Beaconing Sequences (BS1 and BS2) are defined to allow for a means to determine far end presence and readiness capability for initiating link training. These sequences can influence the power subsystem of low power nodes, if so enabled. This entails the following functions:
For the BS1, a low frequency (1 KHz) waveform on data bit 0 that can be detected under both Aux and Primary Power to indicate that far end is there and desires to begin training. For the BS2, a low frequency (1 KHz) waveform on data bit 0 that can be detected under both Aux and Primary Power. It can only be sent when the local end is on Primary power.
The FIO link encoding layer provides two link training sequence (TS1 and TS2). Both training sequences are eight symbol ordered set are chosen to optimize the training process. Both training sequences are 80-bits (64 ns) long, contain a unique code group (comma) for byte and word synchronization, and have a maximum number of signal transitions to facilitate PLL locking and bit de-skew. The two training sequences are used for:
Training sequences are the first code groups transmitted on a link. The training sequences ensure deterministic initialization of both transmitter and receiver initial state.
It may be necessary to reestablish flit synchronization as part of link layer error recovery. When requested, the link encoding layer can transmit a link training sequence two (TS2) to initialize link running disparity and force flit alignment.
Simple flit retransmission is not capable of recovering from all possible link error. The upper layers of the link protocols can force a link clear (complete clearing of all state) by forcing link training sequence one (TS1). Reception of TS1 forces the complete link retraining.
The Link Layer on two chips may be clocked by independent oscillators, with a small frequency differences between the two sources expected. Without compensation, this frequency error would cause link data under-run (low frequency to high frequency) and link data over-run (high frequency to low frequency). Data under-run is easily tolerated without loss of data. Data over-run is unacceptable due to the loss of data. To compensate for the expected frequency error between two chips, SKIPs are inserted in the link layer data flow every 2000 symbol times. This SKIP insertion rate can compensate for total clock frequency difference of 500 ppm or less.
There are a number of “link states”, including:
The LDisabled state indicates this end of the link is disabled from sending or receiving any operations. The FisxOw signals shall be quiesced; the FIOwxIw signals shall be ignored. Entry into the LDisabled state is done through a management agent setting LCSOperationStatePending to the ‘Disabled’ as defined in Table 3 Link State Summary. It is the management agent's responsibility to ensure that traffic is appropriately quiesced if no data loss is required. LDisabled is maintained while LCSDisabled is set to ‘1’. Exiting from the LDisabled state will cause the link to proceed to the state indicated by the value in LCSEnable-Default facility. The valid options are:
The LStandby state indicates that the link is operating on Auxiliary power and thus can only detect beaconing sequences (BS1, BS2). No drivers are active in this state. Receiving either BS1 or BS2 in this state will cause the assertion of the FmsxInt signal if enabled by a LCS bit such to allow for link activity to power-on the chassis. This assertion can be used by the chassis power subsystem to enable the Primary Power. The LStandby is entered by the following process:
The LLSBeacon state indicates that the link is looking for a far end but is only capable of receiving BS1 or BS2 responses as the high speed detection circuitry (implemented in PLLs) is disabled This capability is indicated to any present far end by sending BS1. Only FIOsxIw(0) is active; the other drivers are quiesced. This state is used by switches to continually beacon for attachments while maintaining low power consumption. If the far end is on Auxiliary power and is in Standby, BS1 causes a notification to the power subsystem. LLSBeacon is entered due to following conditions:
The LHSBeacon state indicates that the link is looking for a far end but is capable of receiving BS1, BS2 or TS1 responses as the high speed detection circuitry (implemented in PLLs) is enabled. This capability is indicated to any present far end by sending BS2 Only FIOsxIw(0) is active; the other drivers are quiesced. This state is used by any node wishing to make an attachment at either low or high speed; the attachment speed is thus determined by the present capabilities of the far end. If the far end is on Auxiliary power and is in LStandby) BS2 will cause a notification to the power subsystem for wake-up. LHSBeacon is entered due to following conditions:
TS1 is received. As the high speed circuitry is enabled in this state, this condition leads to the movement to L Ready state so that the near end can commence with the link training sequence.
The values for beacon and suspend should be driven by optical considerations.
The LSleep state indicates that this end of the link is powered and waiting to receive a beaconing sequence (BS1 or BS2) but is not itself sending a beacon. This state is intended to be used by switches to wait on hot plug events or as a means to put the link into a low power state. For the case of preparing a port for hot plug, it can be assumed that the state of the far end is inconsequential; thus, no synchronization is necessary. The steps to enter this state are:
The LReady state indicates this end of the link is high speed enabled and is looking to achieve both frequency lock and flit alignment. Frequency lock will be achieved through any data stream received at the established frequency. This state will exist once the alignment is achieved through the detection of either TS1 or TS2. The difference between these is that TS1 will cause the de-skew function to be recalibrated; TS2 will not cause this recalibration.
Link State Summary
Table 4 is the beginnings of the matrix of the states possible at both ends and the wires between them. Red shading indicated and invalid or error state. Blue indicates a valid “holding state”. Green indicates the normal training and then data passing states. “Con” means “connection”; “NC” means “No Connection”. The Rows of “State” is meant to be the primary end of interest (i.e. what will that ends port logic see such to make decisions on) and also include a “Recv” and “Send” sub-row. “Recv”, thus, has the combination of what the far might be trying to put on the wires base on its local knowledge (Con or NC) and the real connect at both end will produce at the near end. “-” means that the interface is “Null”; we could define what is. TS0, TS1, TS2 and Data are from above or self-defining.
Link State Matrix
The Network Layer.
To implement a multi-function end-station, e.g., a single adapter which implements multiple distinct I/O functions such as multi-port SCSI or a GbE and Fibre Channel combination adapter, an IHV may choose to embed a switch element ASIC on-board—conceptually, this is similar to a PCI bridge (illustrated in
Future I/O routes data via “frames.” A frame is a Future I/O fabric unit of work with the following high-level attributes:
There are two types of frame routing: within a subnet and between subnets. Within a subnet, Future I/O frames are routed between endnodes via switches—switches may be internal to an endnode (e.g., a platform or an expander box) or external (similar in concept to a GbE or Fiber Channel switch).
Local identifiers are 16-bit, endnode identifiers with the following properties:
A subnet is a collection of endnodes and routing elements which use the same address prefix. A Future I/O subnet has the following attributes:
Future I/O must be able to operate without requiring a fabric manager to be up and operational. As such, Future I/O defines a default subnet which is used during endnode and fabric initialization. This default subnet will be used for all routing operations until an optional fabric manager decides to perform a re-con-figuration operation. For a simple Future I/O topology, e.g. a platform with a Future I/O backplane, a fabric manager may never exist or if it does, it may choose to not configure a non-default subnet address. As such, all Future I/O management and routing operations which involve the subnet address, shall be able to operate using the default address. If a fabric manager is operating (e.g., an operating system on a platform is executing), it may reconfigure the default subnet address to one set by a system administrator or returned from some form of directory service. Future I/O does not define how this middleware operates; it defines a set of management messages which may be issued to perform the endnode configuration as appropriate. * If a fabric manager is not present, the autonomous discovery protocol will assign DLIDs. This protocol is described in further below in connection with configuration and discovery.
Basic operation principles for subnet masks include:
Access control is important because an endnode must be protected against unauthorized access at multiple levels—application, kernel, hard-ware, etc. Future I/O specifies two types of routing protection: source route restrictions and hardware fire-walls. Other type may be implemented in Future I/O switches and routers. Note: Additional levels of protection are provided via other services (e.g. Partitioning) within Future I/O or by OSV/ISV/IHV middleware with the later being outside the scope of the Future I/O specification.
Source route restrictions are best implemented in the switch where the source endnode attaches to the Future I/O fabric. Future I/O will define the management messages needed to configure source route restrictions (default is unlimited access within a subnet or between subnets) and it is up to the IHV to determine the switch implementation. Note: While this section focuses on switches, there is nothing to restrict an adapter or SANIC from providing a similar protection mechanism to protect itself from unauthorized access. One possible example implementation is to use a small number of access control bits which are associated with each switch input port (this limits the switch resource requirements to the number of ports*the number of access control bits). The following table illustrates how a set of bits might be interpreted. An alternative implementation which is more robust at the expense of additional resources and complexity (additional management messages and possibly for global headers, the storage and mapping of source Ipv6 addresses), is to provide a set of access control bits or partition key per DLID. This would allow a switch to provide more fine-grain access control on a per source/destination tuple or application partition basis.
A hardware firewall may be constructed by limiting which endnodes may route to another or across sub-nets. For example, a router may be configured to restrict access to a given subnet or individual endnode. This may be accomplished using a subnet mask or by configuring individual source (protocol dependent) addresses which may access this subnet or route to or from a given node within a subnet. A switch may provide similar functionality by expanding its route table to include an additional source/destination access rights table—this functionality is left to the IHV to design and implement as a value-add feature.
Access Control Values
Frame Route Access Allowed
No access - the sender may not route any frames through this port.
The sender is allowed to issue management enumeration frames
and to perform base discovery operations.
The sender is allowed to issue management control messages,
e.g. update the switch/router route tables, reset the switch, etc.
The sender may route application data and connection management
Future I/O allows arbitrarily large fabric topologies to be created, i.e., it is highly scalable. Given the architecture's life span and the growth in the number of elements which may be addressed, Future I/O will use the IPv6 addressing standard as follows:
Future I/O defines two route headers referred to as Local Route Header (“LRH”) and Global Route Header (“GRH”)—all Future I/O components—endnodes, adapters, devices, switches, and routers—shall support both route header formats. Route header selection is determined as follows:
The local route header is a compact, efficient, 8-byte header containing the following four 16-bit fields (one might view it as a compressed Ipv6 header optimized for use within a single Future I/O subnet): The LRH attributes are:
Local Route Header
LRH Efficiency Relative to Data Payload
Data Payload (Bytes)
LRH Efficiency (%)
Future I/O uses the Ipv6 header for global routing of FIO transport headers and optional data. FIO will work with the IETF to obtain the associated Ipv6 Next Header identifiers to uniquely identify the FIO transport headers. See RFC 2460 Internet Protocol, Version 6 (Ipv6) Specification for additional details on how Ipv6 operates. Unless otherwise indicated within this specification, FIO will comply with the RFC 2460 for header definition and addressing. Clarifications:
Future I/O leverages IP technology and concepts to provide a number of customer value propositions. The following provides a sample of how customers directly benefit by leveraging IP technology:
Table 8 illustrates that the local and global routing headers are relatively comparable across most payload types if the applications used data size exchanges were completely uniform. However, most application work loads are a mix between small and large data exchanges with many applications using 80% small and 20% large data exchanges. When this is taken into account, the efficiency delta between the two header formats is sufficiently small to quell any concerns about the GRH size.
GRH Efficiency Relative to Data Payload
Data Payload (Bytes)
GRH Efficiency (%)
Each 16-bit field in an LRH field should be treated as a little endian field to determine bit position within the two-byte field. More particularly, the 16-bit header control field contains control bits used for routing a frame. The destination local identifier (“DLID”) field specifies the local identifier for the destination endnode within a subnet. If the final destination is outside of the subnet, the DLID shall be a router which is routing the frame onwards to its final destination—the LRH is stripped in such a situation. The frame length indicates the length of the frame which is the sum of the route (local and/or global) header length, the transport header length, and the optional payload data length. To calculate the actual payload data length, the route and transport header lengths are subtracted from the frame payload length. Future I/O specifies this frame length encoding scheme to allow frames to be routed while treating the payload (transport and optional payload data) as an opaque sequence of bytes. Additional guidelines:
Header Control Definition
Indicates Future I/O Version. Initial value is set to 0x8.
Subsequent versions of Future I/O are required to be backward
compatible to allow endnodes developed for an earlier version to
inter-operate without modification. However, new functionality
tied to a given version may not be accessible by these earlier
Field indicates if the operation is sequenced (set to 0) or
non-sequenced (set to 1). A sequenced opera-tion is one in
which the frame sequence number is incremented for each
unique frame sent between destination or QP endpoints.
The sequence number is used to maintain strong ordering. If a
missing sequence number is detected (either due to an end-to-
end time-out or due to the arrival of a frame with a sequence
number which exceeds the expected sequence number), the
destination endnode gener-atesa NAK which indicates how the
sender should recover from this error. A non-sequenced
operation is one in which the frame sequence number is not
incremented for each frame sent and ordering is not preserved.
This allows frames to be automatically striped across redun-dant
links within the fabric. A non-sequenced operation should not
be followed by a sequenced opera-tion until all frames from the
first operation have been received and acknowledged.
Next Header - Indicates the header type which follows the LRH,
e.g. FIO transport header. FIO will leverage RFC 2460 for next
header field value. For non-FIO protocols, a SANIC may be
implemented such that the Next Header field acts as an index
into a table which contains a list of locally managed QPs. This
allows an endnode hardware to directly target the upper layer
protocol without having to parse the header - this translates
into improved protocol implementation efficiency and scaling
Future I/O provides transparent support for legacy protocols, e.g., IP, IPX, SNA, etc. within a FIO subnet as follows:
Future I/O will support two forms of multicast—unreliable and reliable. Multicast routing and management leverages the existing IP-based multicast standards and algorithms. Future I/O will provide the following:
Applications address endnodes and routing elements using their associated Ipv6 address and are unaware of the corresponding Ipv6-to-DLID mapping. As such, applications will use the Ipv6 Multicast standard (RFC 2375—Ipv6 Multicast Address Assignments) to manage multicast addresses. The OSV, ISV, and IHVs shall provide the associated Ipv6-to-DLID mapping services. Multicast uses the same route header DLID field as a unicast DLID—the difference is how the field is interpreted by the source, destination, and intermediate routing elements.
FIO unreliable multicast is implemented within switches and routers by sending a frame to a DLID which has been previously configured to represent a multicast group.
Reliable multicast is implemented as a series of acknowledged unicasts by the source endnode. For Future I/O version 1.0, reliable multicast is implemented via the Future I/O software verbs as a series of individual sequenced message sends (this functionality may be rolled into hardware in a later version or implementation of Future I/O). In other words, switches and routers view a reliable multicast operation as a normal unicast operation; their route tables are not modified to indicate the associated DLID is participating in a reliable multicast group. A reliable multicast group is set up using the same high-level multicast group operations but the underlying interpretation and actions taken are very different. For example, a join operation requires each participating endnode to know the exact group composition, i.e., it must maintain a complete list of destination addresses and establish reliable datagram end-to-end contexts to transmit data, and to track and generate acknowledgments for each frame received. Additional guidelines:
Striping, i.e., the dynamic sending of frames within a single message to an endnode's QP through multiple fabric paths (nominally using a least busy port algorithm), is used to increase aggregate bandwidth and to reduce the end-to-end message latency. Striping also allows replicated fabric resources (typically deployed to provide fault tolerant fabric paths between endnodes) to their maximum benefit. Future I/O supports striping for two configurations:
Policy based routing (sometimes referred to as static striping) is focused on end-to-end aggregation across multiple QPs to provide one level of fabric and QP QoS—this differs from striping which is focused on end-to-end aggregation to a specific QP. This is illustrated in
A router is an endnode which is used to join Future I/O subnets to Future I/O subnets or to disparate fabric types. Router requirements include:
Router fail-over. A strong RAS solution provides support for replicated ports and fabric instances which allows an application to transparently fail-over or explicitly migrate (e.g., Service Level Objective load-balancing) from one instance to another. To facilitate this, it is strongly recommended that the SANIC driver maintain alternative DLIDs for reliable transport services which can be transparently updated within the SANIC's QP state structure during the fail-over/migration process. Note: At a minimum, each port shall be assigned one LID. If there exist multiple paths within the fabric to a port, these LIDs may be optionally determined and maintained within the endnode for fail-over/migration purposes. Future I/O fabric management (not the autonomous discovery protocol) is responsible for determining this data.
The Transport Layer Protocol.
The description of the network and routing layer found immediately above describes how a frame, with its network header, is sent through the network to the appropriate endnode. Encapsulated within the network frame is the transport header and the data payload. The transport header has the information needed by the endnode to deliver the data payload to the appropriate process or thread residing on the endnode. FIO fabric switches and routers ignore the transport header; they only need the network header to rout the frame to the destination.
User of kernel mode clients of FIO interact with the transport layer indirectly through the Verbs and SANIC provider software layer. These software layers provide a series of calls to manipulate a “queue pair” (“QP”) made up of a send work queue and a receive work queue. The client posts buffers or commands to these queues and hardware transfers data from or into the buffers.
When a QP is created it is associated with one of four transport services. The transport service describes the degree of reliability and to what and how the QP transfers data. An endnode offers four types of transport services:
FIO Transport Services Comparison
M 2 *N QPs
M QPs required
M QPs required
1 QP required
processes on N
on each end
processes on all
Reliable. Any errors not
recovered by flit retransmission
The message is
While in the
(e.g. a failed cable or
delivered in its
switch/router) are handled by the
entirety or not
The message is
SANIC driver without client
at all. Flit
delivered in its
involvement. All messages are
entirety or not
received correctly, without
on errors is
at all. When
duplicates, unordered or dropped
routed out- side
frames. The user client sees
correct data delivery. The QP is
halted only if the destination is
frames, out- of-
corruption is a
dead or all communication paths
order, and stale
function of the
zero copy Data
local QP to one
to the client,
and only one
but the SANIC
remote QP. No
driver shall first
each SANIC it
The reliable connection and reliable datagram services use frame acknowledgment messages called “Ax” as well as sequence numbers in the transport header. These ensure the delivery of every frame, prevent duplicate frames, and allow the detection of out of sequence or missing frames. Once detected, missing frames are retransmitted by the requestor.
The reliable transport services assume incoming frames destined for a particular QP are in order. The network, except for brief periods when it reconfigures itself, will deliver frames in order. Out of order frames are essentially interpreted as missing frames. The endnodes are not expected to reorder incoming out of order frames.
For the reliable services, every request frame has a corresponding Ack frame. While the FIO Architecture Specification does not imply an implementation, it is typical for hardware to generate Ack frames. The far less common case of error recovery whereby frames need to be retransmitted can involve SANIC driver software to schedule the retransmitted frames.
NAK, or “negative acknowledgement,” frames are used to respond to a failed request. NAKs are intended for error handling or error reporting when informing the requestor can help solve the problem or provide debugging insight. For example, a NAK is generated for a sequence error detected on a reliable connection QP. The NAK tells the requestor to resend earlier frames. A NAK is not generated for things like checksum errors or a routing element which delivers the frame to the incorrect destination. Such errors were not the fault of the requestor nor can the requestor ameliorate the problem.
FIO data structures and wire format described in this and related documents follow the following rules:
Bit, Byte, Word, Dword, Qword formats
There are 10 different transport header formats used by the four transport services. Table 14 below lists which frame formats are used to support each of the transport services. The formats of the various transport headers are defined below. The Opcode field in the first byte of the transport header determines the transport frame type.
Frame Formats Used by the Transport Services
by FIO. The
field in the
aAn “X” indicates the transport service can use that particular frame type.
Transport header for “Restart”, “Resync”, or “Inquire” Operations
Sequence number (24)
Source QP# (20)
Destination QP# (20)
FKEY usage for the above commands is a “Work in progress”
Transport header for “Send” Operations
Sequence number (24)
Source QP# (20)
Destination QP# (20)
the length of the Payload is calculated off the route header frame length.
TABLE Transport header for “General Ack” Operations bits bytes 63-56 55-48 47-40 39-32 31-24 23-16 15-8 7-0 0 Opcode (8) Sequence number (24) Original Source QP# (20) TSN (12).
Note The TSN field is copied from the same positions of the request frame, the length of the Payload is calculated off of the route header frame length, the length of the Payload is calculated off of the route header frame length, and, for Acks, the SLIDE field is encoded into the Opcode field. This was done to keep the Ack header small (it will only use a single 16 byte flit).
The General Ack frame is sent in response to a Send, FetchOp request, FetchOp Response, or RDMA Write request that is received at a QP offering one of the two reliable transport services. Note that there is only one Ack opcode available for these commands, so the Ack must be associated with the request frame by use of the QP, TSN, and sequence number, to determine the original request opcode. The sequence number, TSN, and Source QP fields in the Ack frame are replicated from the associated Send or RDMA write request frame. Acks are always sent to the “source ID” stored in the QP or EE context, and matched against the “Source ID” found in the incoming request frame. This is even if the other end's ID for the connection is different. An Ack cannot be sent until the QP context is examined and Various checks are made; including whether to send an Ack or not (unacknowledged service).
Transport header for a NAK Opcode bits bytes 63–56 55–48 47–40 39–32 31–24 23–16 15–8 7–0 0 Opcode (8) Original Sequence Original Source TSN (12) number (24) QP# (20) 8 Original FKEY (32) Original SLIDE Rsvd Destination QP# (20) 16 NAK Code Reserved Note The TSN field is copied from the same positions of the request frame. Transport header for an RDMA write Operation bits bytes 63–56 55–48 47–40 39–32 31–24 23–16 15–8 7–0 0 Opcode (8) Sequence number (24) Source QP# (20) TSN (12) 8 FKEY (32) Destination SLIDE Rsvd QP# (20) 16 Virtual Address 24 R_Key Payload Data 32 Payload Data Note: the length of the Payload is calculated off of the route header frame length. Transport header for an RDMA Read Operation bits bytes 63–56 55–48 47–40 39–32 31–24 23–16 15–8 7–0 0 Opcode (8) Sequence number (24) Source QP# (20) TSN (12) 8 FKEY (32) Destination SLIDE Rsvd QP# (20) 16 Virtual Address 24 R_Key Reserved RDMA Length Transport header for an RDMA Read Ack Operation bits bytes 63–56 55–48 47–40 39–32 31–24 23–16 15–8 7–0 0 Opcode (8) Sequence number (24) Original Source TSN (12) QP# (20) 8 FKEY (32) Original SLIDE Rsvd Destination QP# (20) 16 Payload Data Note: the length of the Payload is calculated off of the route header frame length. Transport header for a FetchOp Command Operation bits bytes 63–56 55–48 47–40 39–32 31–24 23–16 15–8 7–0 0 Opcode (8) Sequence number (24) Source QP# (20) TSN (12) 8 FKEY (32) Destination SLIDE Rsvd QP# (20) 16 Virtual Address 24 R_Key Reserved 32 Swap Data 40 CMP Data (Used for Compare and swap only) Transport header for a FetchOp Ack Operation bits bytes 63–56 55–48 47–40 39–32 31–24 23–16 15–8 7–0 0 Opcode (8) Sequence number (24) Original TSN (12) Destination QP# (20) 8 FKEY (32) Original Source SLIDE Rsvd QP# (20) 16 Pre-Swap return Data Transport header for a FetchOp Release Operation bits bytes 63–56 55–48 47–40 39–32 31–24 23–16 15–8 7–0 0 Opcode (8) Sequence number (24) Original TSN (12) Destination QP# (20) 8 FKEY (32) Original Source SLIDE Rsvd QP# (20) 16 Virtual Address 24 R_Key Reserved.
Table 15 presents the definitions of the opcode field.
TABLE 15 Opcode field 00eeeeeea General Ack 10000000 Connected Send (First or intermediate frame) 10000001 Connected Send (last or only frame) 10000010 Connected RDMA Read 10000011 Connected RDMA Read Ack 10000100 Connected RDMA Write 10000101 Connected Restart 10000110 Connected Inquire 10000111 Connected Resync 10001000 Connected Swap 10001001 Connected CmpSwap 10001010 Connected FetchAdd 10001101 Connected FetchOp Ack 10001110 Connected FetchOp Release 10001111 Connected NAK 10100001 Reliable Datagram Send 10100010 Reliable Datagram RDMA Read 10100011 Reliable Datagram RDMA Read Ack 10100100 Reliable Datagram RDMA Write 10100110 Reliable Datagram Inquire 10100111 Reliable Datagram Resync 10101000 Reliable Datagram Swap 10101001 Reliable Datagram CmpSwap 10101010 Reliable Datagram FetchAdd 10101101 Reliable Datagram FetchOp Ack 10101110 Reliable Datagram FetchOp Release 10101111 Reliable Datagram NAK 11000001 UnReliable Datagram Send 1110xxxx Manufacturer Specific Opcodes (16 codes) 11111111 NAK (Unsupported feature). a“eeeeee” is the request frame's SLIDE field
Note that unlisted opcodes are reserved; they should not be generated by senders, they should be ignored and NAKed (11111111b) by receivers. Reserved values of the opcode field and those that are not supported by the device, are NAKed by endpoint devices. The current frame qualifiers “first”, “intermediate”, “last”, and “only” listed above indicate which frame of the message is represented by the opcode. There are separate opcodes for Reliable Connection, and Reliable Datagram to allow the remote SANIC to determine the mechanism for accessing the connection state (QP context or EE context).
With respect to other fields used in the various headers:
Receiver's Interpretation of Incoming Sequence Number
Exp#a − Rx#b = 0
Exp# − Rx# ! = 0
aExpected Sequence Number
bReceived Sequence Number
The receiver sequence
number did not match the
Receiver not ready (“RNR”)
The receiver cannot
perform the operation at
R_Key not recognized by
Access outside of R_Key
Start address not in range
limit due to Virtual address
specified by R_Key
Access outside of R_Key
Access beyond the end of
limit due to frame length
the range specified by the
Access outside of process
Page protection violation at
space (L_Key violation)
Access outside of Rx buffer
Send message length too
limit for Send
long for the receive buffer
Invalid Opcode (either
reserved or not supported)
Unexpected FIO Version
in route header
Invalid QP (QP number not
Destination does not
support this QP number
Source QP not valid for this
For Reliable Connections,
the source QP must match
the programmed value at
Source LID not valid for this
For Reliable Connections,
the source LID must match
the programmed value at
Source IPV6 address not valid
For Reliable Connections,
for this destination QP
the source address must
match the programmed
value at the destination.
Unexpected destination ID
The receiver's DLID did
not match the value
programmed into the
Unexpected destination IPV6
The receiver's Address did
address (route failure!)
not match the value
programmed into the
Frame length shorter than
At least the header arrived
expected (next frame arrives
correctly, so the NAK can
before frame length is
Correct EE context could
not be located for this
*The Payload field is the contents of the message. For multi-frame messages, the sender breaks up the message into frames of the appropriate size, the receiver must re-assemble the frames into a complete message. It is generally assumed that the endpoints will do the break-up and reassembly in the final
Frame Level Details.
The FIO team is currently investigating flow control at the frame level. If used, it must require minimum hardware state and algorithms since it is assumed that it will be combined with the hardware frame scheduler. Currently we are trying to avoid the use of additional “request to send” and “clear to send” messages because we expect these to increase latency and network bandwidth. We also expect that it would also be impractical to have any sort of credit count. Given the large number of endnodes and QPs in an FIO system, any attempt to pre-establish frame credits would be either not scalable or would require very large pre allocated buffers. As a result, we are looking at the use of adaptive rate mechanisms such as have been proposed for use (or used) in the internet.
FIO specifies flow control at the link level and at the message level. The link level flow control is done with flit level credit counts. It is used so switches and endnodes do not drop frames if a resource (e.g. a switch output port) is temporarily busy. Message level flow control is used if a Send message is sent to a QP and there is no receive buffer available. Assuming receive buffers are available, an endnode is expected to process all incoming request frames.
VLs are used for deadlock avoidance by employing different virtual lanes for request and acknowledgment frames. The transport layer shall not cause deadlock in the FIO fabric. Specifically, it shall never cause back pressure on an input because of back pressure it receives from the network. To help meet this requirement, FIO defines that a different virtual lane be used for acknowledgment frames than the request frames. A full description on the use of Virtual lanes is located the discussion on Quality of Service below.
Protection against ghost frames is also provided. Ghost frames, or frames that may be “lost” in the network following an error or network route change, are handled by use of the sequence numbers. See the sections on sequence numbers above and on error handling below Reliable Datagrams and frame level error recovery below for Reliable Connections for more detail.
Frame level timeouts are maintained by the QPs for Reliable Connections and the “end to end” (EE) connections for Reliable Datagrams. These are set by default, to values that work well for FIO subnet use. If the connection goes outside the subnet, the values must be programmed to higher values appropriately. See the discussion on connection management below. Timeouts for the RNR mechanism should be established by software. For Reliable Connection, the timeout is based on waiting a bit longer than the longest expected software response time. When this times out, the FIO consumer is notified. For Reliable Datagram, the timeout is based on a reasonable “poll” interval. The SANIC also maintains a “poll counter” to determine the number of times to poll before the FIO consumer is notified. Both the RNR timeouts must be set to be longer than the frame level timeouts. These timeouts will be not less than the amount specified, but may be considerably more. Implementation note: It is expected that the timeout will use the QP frame scheduler as a timing mechanism. As a result, the timeouts will probably be an integral number of scheduler cycles.
Turning first to the send operation, there is a one to one correspondence between send WQEs on the transmitting side and receive WQEs on the receiving side. The data to be transmitted is given by the “local address” and “data length” parameters given by the calling verb. There may be up to 8 address and length pairs organized into a “gather list” to make up a single message. A zero length is acceptable for the first (only) data region specified, if more than one region is given as a gather list, they must all have non-zero lengths. The SANIC must support a length of a region up to 232−1 bytes. If there is no actual data present in the incoming frames, the receiver should not perform any memory protection checks. Each local data address range is checked for validity against the local page protection table, translated into the appropriate address, read from memory and packed into frames for transmission. Page access violations are reported when the message WQE is dequeued. See the SANIC Memory protection section for more details.
When received, the message is subject to appropriate checks, stored, and each frame is acknowledged. Errors are returned with a NAK, and the appropriate error code. See the discussion on the receive operation immediately below for more details. When the message has been completely sent, and all acknowledgements have been returned to the sender, the WQE is marked “done” for reporting with the appropriate verb.
The receive operation is used to inform the SANIC, the location of buffers used to store incoming data. It also provides a mechanism where the incoming data characteristics (length, immediate data, etc.) can be re-turned to the application.
Unlike other operations, Receives are posted to the receive queue. Each incoming message is stored in the buffers described by a single “receive WQE”. There must be sufficient room in the buffers for the incoming message, or the sender and receiver will get an error. The buffer may be larger than necessary, if so, the state of memory in the buffer beyond the end of the message is left undefined. The buffer which receives the data is given by the “local address” and “data length” parameters given by the calling verb. There may be up to 8 address and length pairs organized into a “scatter list” to accept a single message. A zero length is acceptable for the first (only) data region specified, if more than one region is given as a scatter list, they must all have non-zero lengths. The SANIC must support a length of a region up to 232−1 bytes. As the data comes in, each local data address range is checked for validity against the local page protection table, translated into the appropriate address, and used to store the data. Page access violations are reported when the message WQE is dequeued, as is discussed below relative to memory registration.
The amount of actual data in the Send frames is determined from the route header “Frame Length” field according to the rules for this field.
Receive buffer prefix for Global header mode
Frame length (16)
IP Frame Length
Source Ipv6 Address High Qword (64)
Source Ipv6 Address Low Qword (64)
Destination Ipv6 Address High Qword (64)
Destination Ipv6 Address Low Qword (64)
Source QP# (20)
Beginning of Payload Data
Receive buffer prefix for Local header mode
The RDMA Read operation describes a message that can move a single virtually contiguous block of memory from a remote location to a local location as described by the operation verb. The verb provides a single remote address and its associated protection key and a length of the region to access.
The local buffer which receives the data is given by the “local address” and “data length” parameters given by the calling verb. Scatter lists are not supported for RDMA reads, so only a single address and length pair is allowed. A zero length is acceptable for the data region specified. The SANIC must support a length of a region up to 232−1 bytes. Scatter lists for RDMA read seriously complicate the hardware, primarily due to the out of order properties of the data when striping is used. Since a scatter list can be simulated by multiple RDMA read WQEs, this feature was dropped from FIO. On execution, the SANIC takes the WQE for the operation and breaks up the remote address range into frame sized requests, with each request set up to fill an entire acknowledgment frame. Since RDMA read requests are usually small compared to the response frames, they shall be sent out no faster than the expected incoming data rate, adjusted for the rate control as discussed below in connection with Quality of Service (dos).
The remote endpoint performs the expected protection checks on each frame based on the R_Key, address and size, reads the data from memory, and returns RDMA Read Ack for each frame to the sender. Since the data can be read from the correct memory location regardless of the order of arrival, the sender may mark these operations “non-sequenced”, potentially causing the frames to be delivered out of order. When the frames are marked in this way, the receiver only checks for a constant transport sequence number, (discussed above) for these operations. It is the responsibility of the sender to make certain that an ACK is returned for each request, so that all of the data is transferred.
When the RDMA Read Acknowledge frames arrive, with the data, they are stored according to the local buffer description from the operation verb. When all acknowledges are in, the local WQE is marked with the length transferred (should be the same as the length requested, unless there was an error), and any error information available, and will mark the WQE “done”, ready to be dequeued by the appropriate verb. The sender shall not send any Send or FetchOp commands until all of the Acks for the RDMA Read are back from the receiver. This is required for Reliable Connection service, inherent for Reliable Datagram service. This is required to maintain ordering in the face of a lost RDMA Ack. If an Ack is lost, followed by a Send operation that causes the receiver to invalidate its R_Key, the RDMA Read cannot be retried successfully. No WQEs on the remote node's receive queue are consumed by RDMA operations. No notification is given to the remote node that the request has completed.
Thee RDMA write operation describes a message that can move a block of memory from a local location to a single virtually contiguous remote location as described by the operation verb. The verb provides a single remote address and its associated protection key and a length of the region to access. The local buffer which provides the data is given by the “local address” and “data length” parameters given by the calling verb. There may be up to 8 address and length pairs organized into a “gather list” to accept a single message. A zero length is acceptable for the first (only) data region specified, if more than one region is given as a gather list, they must all have non-zero lengths. The SANIC must support a length of a region up to 232−1 bytes. If there is no actual data specified in the WQE, the sender should not actually place the operation on the wire; the sender should not perform any protection checks, and should complete the WQE immediately.
On execution, the SANIC takes the WQE for the operation and reads frame sized chunks from the local address space, creates the appropriate remote address and length, and creates RDMA write request frames with the data. The remote endpoint performs the expected protection checks on each frame based on the R_Key, address and size, writes the data to memory, and returns RDMA Write Ack for each frame to the sender. Since the data can be written to the correct memory location regardless of the order of arrival, the sender may mark these operations “non-sequenced”, potentially causing the frames to be delivered out of order. When the frames are marked in this way, the receiver only checks for a constant sequence number for these operations (see Section 6.2.5 on page 133 transport sequence numbers). The amount of actual data in the RDMA write frames is determined from the route header “Frame Length” field according to the rules for this field. It is the responsibility of the sender to make certain that an ACK is returned for each request, so that all of the data is transferred. When all acknowledges are in, the local WQE is marked with the length transferred (should be the same as the length requested, unless there was an error), and any error information available, and will mark the WQE “done”, ready to be dequeued by the appropriate verb. No WQEs on the remote node's receive queue are consumed by RDMA operations. No notification is given to the remote node that the request has completed.
Three atomic “FetchOps” operations are supported, Fetch and Add, Compare and Swap if match, and Swap. These are generally expected to be used by applications that require synchronization on some resource that is identified by a memory location. The operations are “atomic” in that they all perform a read and a write at the destination and guarantee that no other devices (Cpus, etc.) are able to slip a write in between these. These are guaranteed to execute at most once. If the FetchOp does not execute on the destination, it is re-ported to the sender. However, like all operations, a non-recoverable error that occurs after execution at the destination, but before the response reaches the sender, is basically unknown as far as the sender is concerned. If this occurs, the connection is lost, and this will ultimately be reported to the application. Any errors that are recoverable will be recovered invisibly to the FIO client.
To enable error recovery, particularly to enable recovery from a lost Ack, FetchOps are handled by a special mechanism that uses a two request/response pair operation. The two request frames are called the “FetchOp Command”, and “FetchOp Release”. A “FetchOp Command” request is executed to the remote SANIC, consisting of the write (and compare for Compare and Swap) data needed for the operation. The remote SANIC either responds to the operation with a FetchOp Ack, or, if currently unable, responds with an RNR NAK. As with other operations, the RNR could be generated due to a memory page fault, or be-cause the SANIC does not have enough resources to process the FetchOp (presumably other FetchOps are being processed).
The “FetchOp Command” request is effectively queued to a special facility at the remote SANIC that processes the FetchOp. This facility shall be implemented with hardware. The “FetchOp Facility” executes the operation, and internally stores the result, the source QP (and EE identifier if in Reliable Datagram mode), the Virtual address and the R_Key to allow error recovery. Following the execution of the operation, the remote SANIC sends a “FetchOp Ack” frame, containing the result data. The sender then returns the data to the FIO consumer as specified in the Verb. The sender then sends a “FetchOp Release” frame to release the remote SANIC's FetchOp resources (the FetchOp result). The Receiver Acks the “FetchOp Release” with a “General Ack”. The WQE is not completed and returned to the FIO consumer until the final acknowledgement arrives.
As with all operations, errors could occur on any of the transfers. If the original “FetchOp Command” re-quest is lost, it is retried using the normal retry procedures. If the retry fails, it is not certain whether the FetchOp took place at the destination, but the QP will be in the error state and the connection broken. (Note that, if the FetchOp is executed on a QP in the Reliable Datagram mode, the failure will result in the EE connection being in the Error state.) If the “FetchOp Release” has an error, the sender will retry as with any operation. If all retries fail, that implies that the connection is lost, and the error recovery routines in the remote SANIC driver will release the FetchOp Facilities' resources. Note: Suggest that only one target be present in a “cache line” of the target node CPU architecture to pre-vent cache line contention. The Size of the operation is always 64-bits. The target must be naturally aligned (low 3 bits of the virtual address must be zero). An error will be reported if the R_Key range does not fully enclose the target. If this or other protection error occurs, it will be reported but will not result in taking an of the “FetchOp Facility” resources.
On getting a FetchOp command, the receiver FetchOp facility checks to see if this is a repeated command by matching its resources against the request parameters (Source QP, EE identification [for Reliable Datagram], R_Key, and VA). If not, the receiver determines whether it has the resources (at least one not busy) to perform the FetchOp. If these resources are not available, the receiver returns an RNR NAK. It is expected that SANICs will have a limited ability to maintain the state required between the “FetchOp Command” and the “FetchOp Release”. Since these operations must be separated by at least one round trip fabric latency, plus any other latencies at the sender, each “resource” is busy for a potentially significant time. It is strongly encouraged that the SANIC be designed to handle a large number of these in order to limit the number of times the RNR protocol will be invoked. If the facility has resources to execute the FetchOps, then the receiver QP performs the memory protection checks as usual. Again the receiver NAKs on any errors.
The receiver's FetchOp Facility then performs the command and stores the request parameters and the result. The facility sets the state of the result re-source to “Busy”, and returns the FetchOp Ack. When the receiver gets the “FetchOp Release” frame, it locates the appropriate resource by matching against the request parameters (Source QP, EE identification [for Reliable Datagram], R_Key, and VA) and releases them for reuse. If the receiver gets a repeated command, as determined by a match between the request parameters and the stored resources, it shall simply return the stored result data with the “FetchOp Ack” frame. If a connection is lost or killed at the receiver, the SANIC driver must examine the “FetchOp Facility” re-sources to determine if any of them should be released.
Initial FetchOp command processing is done for sender FetchOp Actions just as with all other operations. The difference begins when the “FetchOp Ack” frame arrives. When this occurs, the sender must send a “FetchOp release” frame as soon as possible. Implementation note: If request frames are only sent at the QP's scheduled time, it is important to adjust the schedule to send the “FetchOp Release” as quickly as possible. Extra latency at this point delays the sender from getting its result, and causes the receiver's resources to be tied up longer, potentially causing resource contention at the receiver. When the “FetchOp Release” Ack arrives, the WQE should be retired. The sender shall not send any Send or FetchOp commands until the “FetchOp Ack” is back from the receiver. This is required for Reliable Connection service, inherent for Reliable Datagram service. This is required to maintain ordering in the face of a lost “FetchOp Ack”. If an Ack is lost, followed by a Send operation that causes the receiver to invalidate its R_Key, the FetchOp cannot be retried successfully.
The FetchAdd operation provides a way to atomically perform a read, add a constant, and write to a 64-bits, naturally aligned memory region on a remote node. The value initially read is returned to the local endpoint. On the sending end, there is one value and a remote buffer address in the WQE. Local WQE contains the value (Data) to add to the remote buffer at the beginning of the operation. After the operation the remote buffer contains the “original remote value”+“original local value” After the operation the local receive buffer contains the original remote value.
The CmpSwap operation provides a way to atomically perform a read, compare with a constant, and, if the comparison was equal, write a second constant to a 4-bits, naturally aligned memory region on a remote node. The value initially read is returned to the local endpoint. On the sending end, there are two values and remote buffer addresses in the WQE. First value (CmpData) contains the value to compare with the remote buffer. Second value (SwapData) contains the value that is swapped with the remote buffer only if the compare operation (done first) results in equality. After the successful operation the remote buffer contains the original (SwapData) local value. After the successful operation the (SwapData) receive buffer contains the original remote value (which must be the same as the (CmpData) from the WQE). After unsuccessful operation, the (SwapData) receive buffer contains the original remote value (which must be different from the (CmpData) from the WQE).
The Swap operation provides a way to atomically perform a read and write to a 64-bits, naturally aligned memory region on a remote node. The value initially read is returned to the local endpoint. On the sending end, there is one value and a remote buffer address in the WQE. The local WQE value (SwapData) contains the value to swap with the remote buffer at the beginning of the operation. After the operation the remote buffer contains the original local value After the operation the receive buffer (SwapData) contains the original remote value.
These Bind R_Key and Unbind R_Key commands are implemented by the SANIC so that the OS kernal does not have to be involved. This is expected to be implemented by hardware so that the Verb posted on the WQE will be executed in the correct order. The operation does not actually transmit any frames for this command, it initializes tables used by the hardware instead.
The Delay command provides a way for software to implement delay operations in between other WQEs. This is expected to be implemented by hardware so that the Verb posted on the WQE will be executed in the correct order. The operation will use the SANIC scheduler for timing so it will provide delays that are not less than the one specified, but may be more, depending on scheduler settings and traffic.
A QP consists of a pair of Work Queues: a send queue and a receive queue. FIO Consumers post requests, in the form of WQEs (Work Queue Elements), on the Work Queues to send or receive data. A WQE is a memory structure that contains all of the information that the FIO Provider needs to process the request, such as pointers to data buffers. SANICs asynchronously process the posted WQEs and report the results when completed. Each Work Queue has an (implementation dependent) associated mechanism (such as a Doorbell) that is used to notify the SANIC that a new WQE has been posted to a Work Queue. This mechanism (Doorbell) is directly implemented by the SANIC and requires no OS intervention to operate.
Each QP has establishment parameters, these must be fully set prior to operation of the QP, some may be changed with appropriate care during operations. For QPs in the Reliable Datagram mode, the EE context must also be set up prior to any communication, as is discussed more fully in the discussion on connection setup management frames (and especially on connection management control frame definition) below. When setting up a connection, a particular QP service type is requested:
QP and EE Establishment Parameters
Element required for operation
Global/Local header mode
QP + EE
Transmit “hop limit”
for Global header mode
Transmit “Flow Label”
for Global header mode
Transmit “Tclass ”
for Global header mode
(x, y, . . . ): port, DLID,
D_IPV6 address, VL
current path (x, y, . . . ) per port
Frame Ack timeout
Send “Request” sequence
Rx sequence number
Send queue Automatic retry
count for frame errors
Send queue Automatic alternate
on link lost
Send queue timeout for
“Receiver not ready”
Send queue stop/continue on
frame transmission error
Send queue stop/continue on
WQE or protection error
Receive queue stop/continue on
frame transmission error
Receive queue stop/continue on
Receive queue use RNR
protocol for overrun error
Receive queue stop/continue on
WQE or protection error
a(QP) This is part of the QP context state and must be set up by the QP connect or Open operation.
b(EE) This is part of the EE (end to end) context state and must be set up in a (hidden) end to end open operation.
c(D) This is provided by the WQE verb for each frame sent.
dThe connection state is split between the QP and EE context.
e(NA) Not applicable
The following items are controlled locally at the endpoint:
Reliable Connection QPs. The Reliable Connection service uses two QPs that are “connected” such that the contents of send buffers placed on one QP are transported across the SAN and are placed in the buffers described by WQEs on the receive queue of the second QP. This data transfer is reliable in that any error or disruption in communication is handled by the SANIC driver which deals with retransmission and possible use of alternate paths. The client of the Reliable Connection service does not deal with any of the error handling issues. For Global headers, incoming Reliable Connections are identified by the presence of the “FIO” in both the Local and Global next header fields and the Reliable Connection Opcodes in the transport header. For Local only headers, incoming Reliable Connections are identified by the presence of the “FIO” in the “Local next header” field in the LRH, and the Reliable Connection Opcode in the transport header.
A Reliable Connection QP guarantees that all data submitted for transfer will arrive at its destination exactly once, intact, and in the order submitted, in the absence of errors. A WQE is completed with a successful status only when the data have been delivered into the target memory location. If an error occurs that prevents the successful in-order, intact, exactly once delivery of the data into the target memory, the error is reported through the WQE completion mechanism. The Provider guarantees that, when an error occurs, no subsequent WQEs are processed after the WQE that caused the error. Transport errors are considered catastrophic and should be extremely rare for FIO Providers using this service.
Depending on the error handling options selected, errors can cause hardware to stop processing, just be reported in a WQE (on one or both ends), or, in the case of network errors, retried. An error may optionally be delivered if a Send is lost because the Receive Queue is empty. An error will be reported to both ends if the WQE at the head of the receive queue is not of sufficient size to contain the data. Errors that occur on the initiating system, such as WQE format errors or local memory protection errors, cause a WQE to be completed with an unsuccessful status with no data sent onto the network. Errors that occur after a message is placed onto the network, such as a transport error, hardware error, lost packet, reception error, or sequencing error, are delivered to the sender and cause a WQE to be completed with an unsuccessful status. Depending on the error, it may be reported at the sending side or both sides.
A Send or RDMA Write WQE is completed with a successful status once the associated data has been successfully received by the remote endpoint. An RDMA Read WQE is completed with a successful status once the requested data has been written to the target buffers on the initiator's system.
There is no guarantee of the state or order of the memory state until the WQE has been marked done. For Send operations, this is true at both ends, the sender and the receiver. For RDMA write operations, the sender must have the message completed before the remote end can depend on the data. For RDMA read operations, the local side data state is not known until the WQE is marked done. If a message is completed in error, the contents of the buffers associated with the message are not known, one or more of the frames associated with the message may have failed. To allow for more efficient transmission across wide area networks, the frame size used for Reliable Connections can be set to values smaller than 4096. The frame size (“MTU,” or “Minimum Transfer Unit”) must be programmable in 64 byte units to allow the senders and receivers to use the same values.
Reliable Connection QP characteristics
Property/Level of Reliability
Corrupt data detected
Data delivered exactly once
Data order guaranteed
Data loss detected
RDMA Read Support
RDMA Write Support
State of Send/RDMA Write when request
Completed on remote end also
State of in-flight Send/RDMA
First one unknown, others not
Write when error occurs
Max Size of Send messages
At least 232–1 bytes.
For a Reliable Connection implementation with a connected QP, the HW looks at state associated with the QP to determine such things as sequence numbers, the location of a buffer, the destination ID and destination QP number, etc. To support the Reliable Connection, the software must create the QP data structure (context). This connection establishes the data structures needed for sequence numbers and QOS parameters to support the reliable communication. Consider, for example, an outbound request, which involves:
This section of the detailed description shall relate certain operational aspects of the FIO Implementation. Namely, this section shall relate:
As was mentioned above, the physical properties of the FIO interconnect support three predominant environments, with bandwidth, distance and cost optimizations appropriate for these environments:
Tables 9–11 illustrate certain concepts surrounding quality of service with respect to an FIO implementation. This description shall now discuss the Future I/O architectural components upon which a Quality of Service (“QoS”) solution may be developed. The discussion is structured generally as follows:
Due to the variety of terminology and concepts surrounding QoS, this description will first describe the base terms and concepts used by Future I/O QoS. More particularly, the following terms shall be discussed:
Service rate types determine the efficiency with which service level objectives (“SLOs”) are met. Applications and I/O drivers/devices require predictable performance to meet their service level objectives, e.g., an application may require N-Units/s of bandwidth with an average end-to-end latency of M u-seconds to provide a user with an acceptable service rate. SLOs are commonly defined as a function of the end-to-end latency, bandwidth, jitter (i.e., data delivery rate variance), and data loss within a fabric. For this discussion, we will consider three end-to-end service rate types:
Fabric congestion causes data to be dropped, e.g., if a router implements a random early detection (“RES”) policy which allows data to be dropped when congestion occurs), reduced bandwidth throughput, in-creased end-to-end latency and jitter, i.e., heavy congestion translates into poor application QoS. Future I/O provides two techniques to deal with congestion. The first technique is the use of a link level (point-to-point) credit scheme which insures that a flit may not be injected down a link unless the receiver has sufficient resources to receive it. The second technique is the use of Virtual Lanes (VLs) which multiplexes multiple logical data streams onto a single physical link. By providing per VL credits and buffer resources, Future I/O increases fabric utilization in the face of congestion by allowing a data stream on a non-block VL, i.e., a VL with available credits, to route data while a blocked VL is delayed awaiting additional credits.
Future I/O separates transport-level QoS from fabric-level QoS since they are managed independently of each other. Transport-level QoS is primarily focused on the QP scheduling algorithm within a SANIC-scheduling algorithms are not specified by Future I/O and thus will vary from SANIC-to-SANIC and from end device-to-end device. For example, an adapter with limited sharing capabilities may decide to implement a simple round-robin QP scheduling algorithm, i.e., the adapter injects schedules a frame on each QP which can transmit data. On the other hand, a SANIC within a host may implement a multi-level QP ser-vicing algorithm with transaction hints for dynamic frame injection control. What Future I/O does provide is:
Fabric-level QoS is primarily focused on:
Future I/O leverages IP concepts and technology where applicable. The following is a brief review of pos-sible points of interaction—this section requires additional thought on how one might best leverage IP's technology.
Future I/O QoS provides a set of management messages which may be used by a middleware or OS pro-vider to configure SANIC and routing element QoS heuristics which are different than the defaults. Future I/O does not specify the algorithms used to determine these heuristics for a given application work load, fabric topology, etc.
Virtual Lanes (VLs) are a link layer construct provide the following high-level services:
General QP and VL base characteristics include:
For Future I/O version 1.0, VLs 0–7 and 15 are valid. VLs 8–14 are reserved and should not be used by any implementation. Use of a version 1.0 reserved VL will generate a hard link error. Note: Subsequent Future I/O specification versions will maintain complete backward compatibility with version 1.0 VL manage-ment messages and service capabilities. VL 0–7 are used for application and general fabric management data transfer. VLs may be individually or as a group assigned a given service rate. For example, one might create two fabric service rates with VLs 0–3 in one group (best effort) and VLs 4–7 (differentiated services) in the other. The service rates and al-gorithms for each group may be adjusted independently. Continuing with this example, the VL 0–3 group executes using a best effort QoS policy. Thus, the service rate heuristic might be to use a simple round-robin technique whereby one flit (if present and sufficient credits exist) on each VL is sent before returning to the initial service VL (this too may be varied by tracking which VL was serviced last within this group and starting with the next VL in this service group). Note: If a differentiated service flit were to arrive, the differentiated service group would be placed in the scheduling queue immediate service upon the current flit completing its transmission on the link. Now, the VL 4–7 group may have a different policy for servicing individual VLs. For example, it may choose to always start with VL 7 when flits are available and then round robin or perform a weighted service algorithm, e.g., N flits on VL 7, N-2 on VL6, etc. For VL 15: 1) Reserved for fabric management operations involved in route table and fabric deadlock. 2) VL 15 operates in a store-and-forward which allows it to proceed in the face of fabric deadlock independent of the routing table contents. 3) VL 15 is not used for normal management messages which may be sent on any of the other available VLs. 4) VL 15 operates as the highest priority VL for scheduling purposes, i.e., regardless of the existing VL scheduling heuristics, if a management message needs to be sent on VL 15, the message will be immediately scheduled once the currently transmitting flit has finished. 5) To prevent denial-of-service attacks via consuming VL 15 bandwidth, only trusted code may issue the associated management messages which execute on VL 15.
Basic characteristics for priority arbitration include:
Arbitration cycles occur when the object counter expires for arbitrating to another service priority level. When a VL at a given priority level is granted the right to transmit, a transmit object counter is preset to Xmit_On, which is configurable. This counter counts out this number while sending objects. When the counter expires, the VL logic begins counting other objects that go by against Xmit_Off. When this counter expires, if there are more objects to transmit at this VL, the VL logic asserts its request to transmit until the Xmit_On counter expires again. If the objects cannot be transmitted or there are no objects to transmit before the Xmit_On counter expires, the VL request is dropped and the Xmit_Off counter is started. If no VL requests are asserted, each Xmit_Off counter is accelerated at the same rate until VL requests are as-serted or the counters expire.
Implementation note: The “acceleration” of the counters may be done in any fashion; one way might be to increment the counters each clock. If the arbitration is pipelined sufficiently ahead of the actual use, this will probably hide the delay. If not, then idles may end up being inserted unnecessarily.
The requests to transmit from the various VLs are prioritized with VL 7 the highest, and the appropriate VL is granted the request. This prioritization is invoked whenever there are more than one VL ready to transmit at one time. The table below shows the arbitration states for each VL and what causes these states to change:
If no more objects,
running If no more
but NO other VL
VL_Grant * On
objects, but another VL
has a request, goto
has a request, goto st3
st4 When counter
expires, goto st2
Waiting for VL grant, no
occurs, goto st0
running and objects
expires, goto st1
!VL_Grant * Off
If an object becomes
running and no more
ready, goto st2; If
!VL_Grant * Off *
objects to transmit at this
no other VL has a
VL, but other VLs
request, goto st4;
expires, goto st5
No VLs requesting, lets
If an object becomes
accelerate the Xmit_Off
ready, goto st2; If
!VL_Grant * Off *
counters in case that's
an other VL has a
what is holding us off
request, goto st3;
expires, goto st5
No counters running,
If an object becomes
nothing to do
ready, goto st1.
Appropriate setting of the VL control registers allows a wide range of choices in arbitration policy. Possible choices are “Strict priority” with VL 6,7 the highest, “Round robin” among VLs, and various band-width allocations to different VLs (which of course only applies when all VLs are contending). “Strict priority” with VL 0,1 the highest is not possible, but can be approached by setting long off times for VL 6,7 and high on times for the VL 0, 1 etc.
Arbitration among the VL0–VL7 data queues has 8 registers, and four control bits to control it. The Xmit_On registers can have values from 0 to 31, where 0 has the meaning “infinite” or stay on the VL until no more transmit objects can be transmitted, or a higher priority VL wins the arbitration. Xmit_Off registers can have values from 1 to 32 where 32 is coded as zero. Table 10 lists some example settings and their bandwidth allocations.
Example Arbitration control register settings
The register layout for these values is shown below in Table 11:
VL Arbitration control registers
Implementation note: The Xmit_On and Xmit_Off registers are described as if they operate on both the re-quest and Ack VLs equally; that is the object level requests for the pair of VLs is ORed into a single request for the purposes of the bandwidth arbitration. When the request is granted the Ack VL is expected to be serviced at the higher priority, sharing the bandwidth with the request VL. If this is not considered accept-able, due to the concern for Ack VLs hogging the bandwidth, then the implementation may use counters for each VL independently and initialize the Xmit_Off values to twice the programmed value.
Programming of the Xmit_On and Xmit_Off registers should use the following formula for best results:
1/(Xmit — Off/Xmit — On+1)+1/(Xmit — Off/Xmit — On+1)+1/(Xmit — Off/Xmit — On+1)+1/(Xmit — Off/Xmit — On+1)=1
If the sum of terms is much greater than one, then the priority of the higher order VLs will be satisfied first, at potential starvation to low order VLs. If the sum is much less than one, then the logic may insert unnecessary idle flits (depending on the effectiveness of the “acceleration” mechanism described above).
In general it is best to use the smallest values of the Xmit_On and Xmit_Off that will satisfy the application needs. For example, use of 2, 2 instead of 1, 1 at a VL will both achieve about 50% bandwidth allocation, but the 2, 2 values will create longer arbitration intervals. The programming of the arbitration control registers is expected to be done to the same values fabric wide; use of different settings in different components will generally reduce performance and QOS guarantees.
Layered Error Management
This section will describe portions of the Future I/O layered error management architecture, or error handling methods. Before delving into the details we will describe the philosophy, and terms, behind the Future I/O Layered Error Management Architecture. Future I/O relies on several error detection mechanisms to determine if a fault exists within a Future I/O fabric component. The error detection mechanisms are performed at each interface layer of the Future I/O Archi-tecture: Verb, Message, Frame, Flit, and Electrical. Once an error has been detected the Future I/O LEMA uses several mechanisms to contain and recover from the error. Finally, the Future I/O LEMA uses several mechanisms to remove the faulty component and replace it. The following section will provide an over-view of Future I/O's Layered Error Management Architecture. The remainder of this chapter will describe the details of Future I/O's LEMA.
Following is a description of the philosophy behind the Future I/O LEMA, as well as the terms used by the Future I/O LEMA. Some of the terms were derived from: Fault Tolerance, Principles and Practice; T. Anderson and P. A. Lee; 1981; Copyright Prentice Hall International, Inc. More particularly:
Excessive Temporary Errors—An excessive temporary errors condition (a.k.a. intermittent error) occurs when the number of temporary errors exceeds the maximum number expected over a given time period. An “Excessive Temporary Errors” condition can result from an intermittent transient fault within a Future I/O component which resurfaces periodically as a temporary error. If the time period of the intermittent transient fault is too small, it is considered a permanent error.
Future I/O supports the following error detection, containment, and recovery mechanisms:
There are three cases for removing a faulty component in the Future I/O fabric:
There are three removal philosophies which may be employed by the Future I/O Consumer process:
All three of these mechanisms are believed to be needed to support all different O/S, plat-form, and I/O types. For instance, a high end configuration may have full redundancy of all fabric components and it's O/S may support an abort operation, so it may choose the second approach. A low-end configuration may have partial redundancy and would require quiesce if the fault is in a resource that is redundant, but is part of a component which has non-re-dundant resources as well.
Similar to component removal there are three cases for activating a replaced, fully operational component in the Future I/O fabric:
All three of these options are believed to be needed to support all the different O/S, platform, and I/O types.
As described above Future I/O employs a layered error management architecture. Faults can occur at each layer in the Future I/O architecture. This chapter summarizes the:
The following sections summarize the errors detected at each layer of the Future I/O architecture. The current contents are not complete.
Interface Description of Layer Type of Check Check error check Verb Interface Communication Communication Immediate down service is not operational. Interface Error resource Error occurred due to insufficient resources. Interface Invalid MTU MTU attribute for a QP was invalid or not supported. Interface Invalid parameter One or more of the input parameters was invalid. Interface Invalid RDMA read A memory or QP attribute requested RDMA read support, but the SANIC provider doesn't support it. Interface Invalid service type Service type attribute for a QP was invalid or not supported. Interface No open No previous open communication communication done request has completed. Interface No WQE done No WQE was ready to be dequeued. Interface Not done No WQEs are maked completed on the specified queue. Interface Out of (remote) Abnormal keys condition indicating the SANIC has ran out of Remote Keys. Interface Rejected invalid Communication service ID service was rejected by the remote node due to an invalid service ID. Interface Rejected QP Communication attributes service was rejected by the remote node due to invalid QP attributes. Interface Rejected resources Communication service was rejected by the remote node due to lack of SANIC resources. Timing Timeout The verb timed out. Verb A- Interface Communication lost synchronous Interface QP overrun Interface Receive queue empty Interface Local protection error Interface Remote protection error Interface Transport error Interface Flit error Interface SANIC error. Interface Communication lost Interface QP overrun Interface Receive queue empty Interface Local protection error Interface Remote protection error Interface Transport error Interface Flit error Interface SANIC error Frame Interface Invalid QP Frame referenced an invalid (or closed) QP (local or remote). Interface Remote Key was Remote key was invalid invalid for the memory region referenced. Interface Sequence lost Receiver (or sender) received a frame (or Ack) with a frame sequence count that is outside of the sequence window for outstanding frames. Interface Invalid frame key Frame key was invalid Interface Transaction Receiver (or sender) sequence lost received a frame (or Ack) with a transaction sequence count that is outside of the sequence window for outstanding frames. Timing Timeout The frame timed out at the receiver. Timing Resource not ready Receiver did not have resources (e.g. WQE) available to receive the frame. Flit Coding One coding check to Link segment CRC cover: check occurred. Coding - Link CRC End to End CRC - End-end CRC check occurred. Coding - Wire coding check Wire coding (e.g. invalid 8b/10b code) check occurred. Coding Excessive An excessive temporary errors number of coding and timeout checks occurred. Interface One interface check Sender attempted to to cover: - No buffer send a flit, but available - Invalid receiver did not VL - Sequence lost have a buffer available (e.g. sender sent flit without flit credits). Interface A flit was received on an unsupported VL. Interface Sender received an ACK with a sequence count that is outside of the sequence window for outstanding flits. Timing One timing check to Link training cover: - Training sequence was sequence failed - attempted and Flit timeout failed. Timing The flit timed out at the sender. Electrical Timing Link dead Link alive is not active.
This section describes some error recovery procedures used at each layer of the Future I/O architecture.
The link layer is responsible for flit level error recovery procedures.
Three types of errors are detected at the flit level: interface checks, coding checks, and timing checks.
The receiver is responsible for maintaining an interface check counter and a timing check counter.
Turning now to
Thus, one particular aspect of the method 6600 is the locking and handshaking in the boxes 6610, 6620 of
The composition of the first and second training sequences will also be implementation specific. For instance, an embodiment implementing an FIO system will use the following training sequences:
In order that those having the benefit of this disclosure may appreciate the invention more fully,
Turning now to
In the illustrated embodiment, the devices are connected via a point to point serial link 6850. The serial link may comprise an even number of lanes or channels through which data is transmitted. Of the even number of lanes, half will transmit serial data in one direction while the other half transmits data in the opposite direction. In the preferred embodiment, the physical links will implement 1, 4, or 12 lanes in each direction. Thus, each link will have a total of 2, 8, or 24 total lanes. Note that alternative implementations may employ alternative kinds of links.
In the latter two implementations (i.e., the 4 and 12 lane links), a single stream of bytes arriving at the input to the physical link are distributed evenly, or “striped”, among the multiple lanes. In the case of the 12-lane link, the first byte is sent to the first lane, the second byte is sent to the second lane and so on until the 12th byte is sent to the 12th lane. At that point, the byte distribution cycles back to the first lane and the process continues. Thus, over time, each lane will carry an equal 1/12th share of the bandwidth that the entire link carries. The same process and technique are used in the 4 lane link. Alternative embodiments with different numbers of lanes would preferably implement this striping process.
Once the bytes are distributed among the individual lanes, the 8-bit words (i.e., bytes) are encoded into 10-bit words, serialized, and transmitted through the physical link. At the output of the physical link, the 10-bit words are decoded back to 8-bit bytes and are re-ordered to form the original stream of 8-bit words.
The North bridge logic 6920 also may provide an interface to an Accelerated Graphics Port (“AGP”) bus that supports a graphics controller 6930 for driving the video display 6940. If the computer system 6810 does not include an AGP bus, the graphics controller 6930 may reside on the host bus.
Various peripheral devices that implement the host bus protocol may reside on the host bus. For example, a modem 6950, and network interface card (“NIC”) 6960 are shown coupled to the host bus in
As discussed above the serial data sent through the physical links is sent in the form of packets. The preferred embodiment employs packetized data and uses specialized packets, more precisely known as ordered sets, that are called Training Sequence 1 (“TS1”) and Training Sequence 2 (“TS2”) to train the system devices prior to transmitting “real” data through the switched system. The actual content and structure of the training sequences are discussed in further detail above, for an FIO implementation, and below, for an InfiniBand implementation.
The training data, TS1 7120 and TS2 7130 are known data that are transmitted between Port A 7100 and Port B 7110. The training sequences perform at least two functions. First, the initiation and duration of the training sequence is established by the transmission and reception of the training sequences. Secondly, given that the training sequences contain pre-determined data, the transmit and receive ports can use this knowledge to correct for any problems (e.g., data inversion, lane order, lane skew) that may be encountered. Since these problems constant and permanent conditions of the physical media, the training sequence may be used to automatically correct the errors for all subsequent data transferred through that physical link.
Note also that the order in which Port A 7100 and Port B7110 configure their receivers is irrelevant in this particular implementation. Thus, regardless of what Port A 7100 is doing, Port B7110 can come out of reset and immediately transmit the TS1 and, upon receipt of another TS1 transmitted by another port (regardless of whether that port is Port A 7100), begin transmitting TS2. Port A 7100 can begin doing the same thing. When both Port A 7100 and Port B 7110 are transmitting TS2, they can then begin transferring information to each other. The process is scalable across an entire computing system comprising many ports.
Meanwhile, Port A 7100 has been receiving TS1 data 7120 from Port B 7110 and performs the same signal integrity checks and correction that Port B has completed. Once both ports are trained with TS1 data 7120, the ports will proceed by sending TS2 training data 7130. This second training sequence serves as a redundancy check to verify that the Ports were trained properly with TS1 data 7120. In addition, the TS2 data 7130 signifies that both ports are trained and are ready to transmit and receive data packets 7170. Once a port is transmitting and receiving the TS2 training sequence, it may begin sending data. With physical link errors corrected by the training sequences, the data packets 7180 can then transmitted and received by the ports as intended.
In the event the training sequence fails, a timeout may occur and the affected port may be powered down or otherwise deactivated. Thus, when a transmission fault occurs, locating the problems in the physical link is facilitated by determining which port has been deactivated. By comparison, failure isolation in a bus architecture can be difficult because if one attached device fails, the entire system may fail. Discovering which device caused the failure is typically a hit-or-miss proposition.
Each word in the training sequence is a 10-bit word that complies with the 8B/10B code discussed above. The first row (COM) in each column is a comma delimiter with a preferred code name K28.5. The second row in each column is a lane identifier that is unique to each lane in the physical link. A table of preferred lane identifiers is shown in
The comma delimiter and lane identifiers are chosen to be insensitive to data inversion. That is, inverting a comma delimiter or a lane identifier symbol changes only the running disparity and not the symbol itself. Consider the 10-bit word for the comma delimiter K28.5. For a negative running disparity, the word is 001111 1010. For a positive running disparity, the word is 110000 0101. These two words are complements of each other. Inverting all the bits in the first word will yield the second word and vice-versa.
Hence, regardless of whether or not a bit inversion has occurred in the physical link, when the receiver port decodes this word, the comma delimiter will result. The same is also true for each of the lane identifiers in
To correct for a crossed differential pair (i.e., bit reversal in a lane), a receiver will decode the 10-bit words arriving in a particular bit stream and determine when the training data starts (as marked by the comma delimiter) and determine the lane number (as marked by the lane identifier). For training sequence 1, the preferred 10-bit code name is D10.2 and the bit sequence for positive running disparity is 010101 0101. The D10.2 code word is chosen for the training sequence because it uses the exact same code word for negative running disparity as it does for positive running disparity. Thus, the receiver expects to receive the 010101 0101 sequence repeated 14 times for each TS1 regardless of the current state of the running disparity. However, if the complementary code word is received (101010 1010), a completely different word is decoded. The inverted word corresponds to the D21.5 code word. If the receiver decodes this inverse word, the decoder will be configured to invert all the subsequent bits received in that particular lane.
The same conditions hold true for training sequence number 2. For training sequence 2, the preferred 10-bit code name is D5.2 and the bit sequence for both positive and negative running disparity is 0101001 0101. The inverted code word (010110 1010) corresponds to code name D26.5. Again, the receiving port will attempt to recover the correct polarity by inverting a bit stream if the D26.5 code words are received. The preferred code names selected for training sequence 1 and training sequence 2 were selected because of their inversion properties. Other code words exhibiting the same properties will also work in alternative embodiments.
If a delay is needed to compensate for advanced clock timing, the retimers 7400, 7410 may insert an additional SKIP word to delay the arrival of subsequent data at the receiving end of the link. This scenario is depicted by the SKIP ordered set 7440 shown at the receiver of Port B 7110. SKIP ordered set 7440 includes two additional SKIP words that have been added by retimer 7400 and retimer 7410. Consequently, a SKIP ordered set that started with three SKIP words now has a total of five SKIP words. Conversely, if an advance is needed to compensate for delayed clock timing, the retimers 7400, 7410 may remove an existing SKIP word to advance the arrival of subsequent data at the receiving end of the link. SKIP ordered set 7430 shows an example of this scenario. SKIP ordered set 7430 contains only one SKIP word as a result of the removal of one SKIP word each by retimer 7400 and retimer 7410. By compensating for clock tolerances, the link and the Ports on either end of the link can operate in a common clock domain.
In the preferred embodiment, the SKIP word name is K28.0 and the associated 10-bit word is 001111 068100 for negative running disparity and 110000 1011 for positive running disparity. As is the case with the COM and lane identifier words, the SKIP word is insensitive to bit inversion. Other code words exhibiting the same property will also work in alternative embodiments.
The adapter 7500 may also convert signals for transmission to a physical link 7520. A byte stream from a local device 7515 is detected and transmitted to a demultiplexer 7590 that stripes bytes from the single byte stream across a number of individual byte streams.
The leftmost block of
The center block of
The rightmost block of
The preferred embodiment discussed above uses a bank of synchronous FIFO register buffers, but other embodiments may work. For instance, the buffers and related logic may be implemented using a field-programmable gate array (“FPGA”) or any other suitable programmable logic device (“PLD”).
To prevent buffer overflow, a time limit is implemented on attempting to locate COM words across each lane of the link. In the preferred embodiment, a time limit of 4 clock times is used. This time limit corresponds to one fourth of the length of the training sequence. If COM words cannot be found across all lanes of the link within four clock times, the buffers are cleared and the de-skew process is restarted. By limiting the time required to de-skew the multiple lanes, the size of the FIFOs may be reduced and the logic required to implement the de-skew is simplified.
In addition to correcting static skew errors as discussed above, the synchronous FIFOs may advantageously provide an additional measure of correction for active skew errors such as jitter and drift. The receiver in each channel in the link is frequency and phase locked to the transmitter, but the active components in the physical link responsible for serial/parallel conversion and the retimers 7400, 710 shown in
Thus, the present invention manifests many advantageous characteristics in its various embodiments and implementations. In an implementation in accordance with the InfiniBand Specification, the present invention provides techniques for:
Note that, in some embodiments, the first and second training sequences may be identical, or match. More particularly, if the ports have an automatic configuration capability for their receivers, the first and second training sequences may match as long as they indicate the true state of the receiver. In this instance, the method 6600 in
This concludes the detailed description. Some portions of the detailed descriptions herein are consequently presented in terms of a software implemented process involving symbolic representations of operations on data bits within a computer memory. These descriptions and representations are the means used by those in the art to most effectively convey the substance of their work to others skilled in the art. The process and operation require physical manipulations of physical quantities. Usually, though not necessarily, these quantities take the form of electrical, magnetic, or optical signals capable of being stored, transferred, combined, compared, and otherwise manipulated. It has proven convenient at times, principally for reasons of common usage, to refer to these signals as bits, values, elements, symbols, characters, terms, numbers, or the like.
It should be borne in mind, however, that all of these and similar terms are to be associated with the appropriate physical quantities and are merely convenient labels applied to these quantifies. Unless specifically stated or otherwise as may be apparent, throughout the present disclosure, these descriptions refer to the action and processes of an electronic device, that manipulates and transforms data represented as physical (electronic, magnetic, or optical) quantities within some electronic device's storage into other data similarly represented as physical quantities within the storage, or in transmission or display devices. Exemplary of the terms denoting such a description are, without limitation, the terms “processing,” “computing,” “calculating,” “determining,” “displaying,” and the like.
Furthermore, the particular embodiments disclosed above are illustrative only, as the invention may be modified and practiced in different but equivalent manners apparent to those skilled in the art having the benefit of the teachings herein. Although the present invention is one aspect of the FIO Architecture Specification, it is not so limited. The invention may be utilized in computing systems that do not employ FIO, as specified or even FIO, as may become apparent to those skilled in the art having the benefit of this disclosure. Furthermore, no limitations are intended to the details of construction or design herein shown, other than as described in the claims below. It is therefore evident that the particular embodiments disclosed above may be altered or modified and all such variations are considered within the scope and spirit of the invention. Accordingly, the protection sought herein is as set forth in the claims below.
The following documents are hereby incorporated by reference as if expressly set forth verbatim herein:
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US4570220 *||Nov 25, 1983||Feb 11, 1986||Intel Corporation||High speed parallel bus and data transfer method|
|US4584684 *||Nov 22, 1983||Apr 22, 1986||Pioneer Electronic Corp.||Data transmission method|
|US4674037 *||Jan 11, 1985||Jun 16, 1987||Hitachi, Ltd.||Data processing system with an enhanced communication control system|
|US5548732 *||Jan 19, 1993||Aug 20, 1996||Hitachi, Ltd.||Bus adapter for transferring signals between a master unit and a slave unit, and system including the bus adapter|
|US5600793 *||Dec 20, 1994||Feb 4, 1997||International Business Machines Corporation||Method and system of bi-directional parallel port data transfer between data processing systems|
|US5835498 *||Jun 14, 1996||Nov 10, 1998||Silicon Image, Inc.||System and method for sending multiple data signals over a serial link|
|US5961616 *||Mar 26, 1997||Oct 5, 1999||Oki Data Corporation||Data transfer system|
|US6202096 *||Apr 14, 1998||Mar 13, 2001||Hewlett-Packard Company||Method and apparatus for device interaction by protocol|
|US6493320 *||Feb 12, 1999||Dec 10, 2002||Fujitsu Limited||Automatic initialization and tuning across a high speed, plesiochronous, parallel link|
|US6493343 *||Dec 30, 1998||Dec 10, 2002||Compaq Information Technologies Group||System and method for implementing multi-pathing data transfers in a system area network|
|US6667993 *||Aug 3, 1999||Dec 23, 2003||Texas Instruments Incorporated||Communication channel|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US7171493 *||Dec 19, 2001||Jan 30, 2007||The Charles Stark Draper Laboratory||Camouflage of network traffic to resist attack|
|US7216188 *||Mar 29, 2003||May 8, 2007||Emc Corporation||Techniques for accessing devices through a set of serial buses automatically setting unique enclosure addresses and detecting non-unique enclosure addresses upon initialization|
|US7240158 *||Jan 21, 2005||Jul 3, 2007||Jp Morgan Chase & Co.||System and method for caching results|
|US7243121||Sep 8, 2005||Jul 10, 2007||Jp Morgan Chase & Co.||System and method for dividing computations|
|US7287061 *||Aug 16, 2002||Oct 23, 2007||Nec Corporation||Data transfer apparatus and data transfer method|
|US7320094 *||Jul 22, 2003||Jan 15, 2008||Intel Corporation||Retraining derived clock receivers|
|US7327749 *||Mar 29, 2004||Feb 5, 2008||Sun Microsystems, Inc.||Combined buffering of infiniband virtual lanes and queue pairs|
|US7334046||Aug 5, 2002||Feb 19, 2008||Qlogic, Corporation||System and method for optimizing frame routing in a network|
|US7334177 *||May 26, 2004||Feb 19, 2008||Motorola, Inc.||Method and system for tracking sequence numbers|
|US7340167||Oct 8, 2004||Mar 4, 2008||Qlogic, Corporation||Fibre channel transparent switch for mixed switch fabrics|
|US7340469 *||Apr 15, 2005||Mar 4, 2008||George Mason Intellectual Properties, Inc.||Implementing security policies in software development tools|
|US7342934 *||Mar 29, 2004||Mar 11, 2008||Sun Microsystems, Inc.||System and method for interleaving infiniband sends and RDMA read responses in a single receive queue|
|US7343403 *||Mar 29, 2001||Mar 11, 2008||Mitsubishi Denki Kabushiki Kaisha||Network management system for remote setup of control equipment in a network|
|US7346064 *||Sep 26, 2002||Mar 18, 2008||Intel Corporation||Routing packets in packet-based input/output communications|
|US7346702 *||Aug 23, 2001||Mar 18, 2008||Voltaire Ltd.||System and method for highly scalable high-speed content-based filtering and load balancing in interconnected fabrics|
|US7352701||Sep 19, 2003||Apr 1, 2008||Qlogic, Corporation||Buffer to buffer credit recovery for in-line fibre channel credit extension devices|
|US7355966||Jul 12, 2004||Apr 8, 2008||Qlogic, Corporation||Method and system for minimizing disruption in common-access networks|
|US7362717||Oct 3, 2002||Apr 22, 2008||Qlogic, Corporation||Method and system for using distributed name servers in multi-module fibre channel switches|
|US7370082 *||May 9, 2003||May 6, 2008||Microsoft Corporation||Remote invalidation of pre-shared RDMA key|
|US7376693||Jun 20, 2002||May 20, 2008||Jp Morgan Chase & Company||System architecture for distributed computing and method of using the system|
|US7379453 *||Mar 29, 2004||May 27, 2008||Sun Microsystems, Inc.||Method and apparatus for transferring multiple packets from hardware|
|US7388843||Jul 12, 2004||Jun 17, 2008||Qlogic, Corporation||Method and apparatus for testing loop pathway integrity in a fibre channel arbitrated loop|
|US7404020||Jul 20, 2004||Jul 22, 2008||Qlogic, Corporation||Integrated fibre channel fabric controller|
|US7406092 *||Jul 20, 2004||Jul 29, 2008||Qlogic, Corporation||Programmable pseudo virtual lanes for fibre channel systems|
|US7409432 *||Oct 19, 2000||Aug 5, 2008||International Business Machines Corporation||Efficient process for handover between subnet managers|
|US7415539 *||Sep 28, 2001||Aug 19, 2008||Siebel Systems, Inc.||Method and apparatus for detecting insufficient memory for data extraction processes|
|US7428598 *||Nov 20, 2003||Sep 23, 2008||International Business Machines Corporation||Infiniband multicast operation in an LPAR environment|
|US7437547||Sep 26, 2007||Oct 14, 2008||Microsoft Corporation||Method and computer program product for offloading processing tasks from software to hardware|
|US7453880 *||May 23, 2002||Nov 18, 2008||Allied Telesis Kabushiki Kaisha||Media converter|
|US7464016 *||Nov 9, 2001||Dec 9, 2008||Sun Microsystems, Inc.||Hot plug and hot pull system simulation|
|US7474661 *||Mar 26, 2004||Jan 6, 2009||Samsung Electronics Co., Ltd.||Apparatus and method for distributing forwarding table lookup operations among a plurality of microengines in a high-speed routing node|
|US7483997 *||Nov 13, 2001||Jan 27, 2009||Mitsubishi Electric Corporation||Data distribution control device and data distribution control method|
|US7486689 *||Mar 29, 2004||Feb 3, 2009||Sun Microsystems, Inc.||System and method for mapping InfiniBand communications to an external port, with combined buffering of virtual lanes and queue pairs|
|US7496698 *||Mar 22, 2005||Feb 24, 2009||International Business Machines Corporation||System and method for efficient implementation of a shared receive queue|
|US7502338 *||Dec 19, 2003||Mar 10, 2009||Apple Inc.||De-emphasis training on a point-to-point connection|
|US7512109 *||Sep 30, 2002||Mar 31, 2009||Intel Corporation||Slot structure for radio communications system|
|US7525983 *||Jul 20, 2004||Apr 28, 2009||Qlogic, Corporation||Method and system for selecting virtual lanes in fibre channel switches|
|US7536473 *||Aug 23, 2002||May 19, 2009||Intel Corporation||General input/output architecture, protocol and related methods to implement flow control|
|US7548560||Feb 27, 2006||Jun 16, 2009||Qlogic, Corporation||Method and system for checking frame-length in fibre channel frames|
|US7555606||Jul 3, 2007||Jun 30, 2009||Jp Morgan Chase & Co.||System and method for caching results|
|US7593936||Aug 11, 2004||Sep 22, 2009||Triumfant, Inc.||Systems and methods for automated computer support|
|US7603476 *||Oct 13, 2009||Sap Ag||Pseudo-synchronous messaging|
|US7606253||Jan 12, 2004||Oct 20, 2009||Hewlett-Packard Development Company, L.P.||Successful transactions|
|US7613864||Nov 3, 2009||Sun Microsystems, Inc.||Device sharing|
|US7613958||Jan 12, 2004||Nov 3, 2009||Hewlett-Packard Development Company, L.P.||Error detection in a system having coupled channels|
|US7620741||Nov 17, 2009||Sun Microsystems, Inc.||Proxy-based device sharing|
|US7624213 *||Nov 24, 2009||Hewlett-Packard Development Company, L.P.||Passing identification information|
|US7630392 *||May 31, 2005||Dec 8, 2009||Cisco Technology, Inc.||Multi-homing using controlled route leakage at a backup service provider|
|US7639616||Jun 8, 2004||Dec 29, 2009||Sun Microsystems, Inc.||Adaptive cut-through algorithm|
|US7640547||Dec 29, 2009||Jpmorgan Chase & Co.||System and method for allocating computing resources of a distributed computing system|
|US7653066 *||Jan 26, 2010||Cisco Technology Inc.||Method and apparatus for guaranteed in-order delivery for FICON over SONET/SDH transport|
|US7653911 *||Jan 26, 2010||Alcatel-Lucent Usa Inc.||Implicit interprocess communications (IPC) versioning support|
|US7672222||Jan 12, 2004||Mar 2, 2010||Hewlett-Packard Development Company, L.P.||Link failures|
|US7673098 *||Mar 2, 2010||Sun Microsystems, Inc.||System and method for synchronizing mirrored and striped disk writes|
|US7680944 *||Mar 16, 2010||Comtrol Corporation||Rapid transport service in a network to peripheral device servers|
|US7684401||Jul 20, 2004||Mar 23, 2010||Qlogic, Corporation||Method and system for using extended fabric features with fibre channel switch elements|
|US7688838 *||May 3, 2006||Mar 30, 2010||Broadcom Corporation||Efficient handling of work requests in a network interface device|
|US7711878 *||May 21, 2004||May 4, 2010||Intel Corporation||Method and apparatus for acknowledgement-based handshake mechanism for interactively training links|
|US7721159||Feb 11, 2005||May 18, 2010||Hewlett-Packard Development Company, L.P.||Passing debug information|
|US7729288||Mar 5, 2007||Jun 1, 2010||Qlogic, Corporation||Zone management in a multi-module fibre channel switch|
|US7733855||Jun 8, 2004||Jun 8, 2010||Oracle America, Inc.||Community separation enforcement|
|US7761529 *||Jul 20, 2010||Intel Corporation||Method, system, and program for managing memory requests by devices|
|US7792115||Jul 20, 2004||Sep 7, 2010||Qlogic, Corporation||Method and system for routing and filtering network data packets in fibre channel systems|
|US7796629 *||Sep 14, 2010||Integrated Device Technology, Inc.||Packet switch with configurable bandwidth allocation precision|
|US7827248||Nov 2, 2010||Randy Oyadomari||Discovery and self-organization of topology in multi-chassis systems|
|US7836352 *||Jun 30, 2006||Nov 16, 2010||Intel Corporation||Method and apparatus for improving high availability in a PCI express link through predictive failure analysis|
|US7843906||Oct 22, 2005||Nov 30, 2010||Habanero Holdings, Inc.||Storage gateway initiator for fabric-backplane enterprise servers|
|US7843907||Oct 22, 2005||Nov 30, 2010||Habanero Holdings, Inc.||Storage gateway target for fabric-backplane enterprise servers|
|US7843954 *||Dec 11, 2008||Nov 30, 2010||Zarlink Semiconductor Inc.||Scaleable VoIP telephone line circuit|
|US7860096||Jun 8, 2004||Dec 28, 2010||Oracle America, Inc.||Switching method and apparatus for use in a communications network|
|US7860097||Feb 12, 2005||Dec 28, 2010||Habanero Holdings, Inc.||Fabric-backplane enterprise servers with VNICs and VLANs|
|US7860961||Feb 12, 2005||Dec 28, 2010||Habanero Holdings, Inc.||Real time notice of new resources for provisioning and management of fabric-backplane enterprise servers|
|US7870444 *||Oct 13, 2005||Jan 11, 2011||Avago Technologies Fiber Ip (Singapore) Pte. Ltd.||System and method for measuring and correcting data lane skews|
|US7873693 *||Jan 18, 2011||Habanero Holdings, Inc.||Multi-chassis fabric-backplane enterprise servers|
|US7873751||Jul 2, 2008||Jan 18, 2011||International Business Machines Corporation||Infiniband multicast operation in an LPAR environment|
|US7894348||Jul 20, 2004||Feb 22, 2011||Qlogic, Corporation||Method and system for congestion control in a fibre channel switch|
|US7894351 *||Feb 22, 2011||International Business Machines Corporation||Detection of link status in blade systems|
|US7907625||Aug 4, 2008||Mar 15, 2011||Integrated Device Technology, Inc.||Power reduction technique for buffered crossbar switch|
|US7908271||Mar 15, 2011||Triumfant, Inc.||System for automated computer support|
|US7925767||Jan 21, 2009||Apr 12, 2011||Canon Kabushiki Kaisha||Device access based on centralized authentication|
|US7930377||Apr 19, 2011||Qlogic, Corporation||Method and system for using boot servers in networks|
|US7934021 *||Apr 26, 2011||Broadcom Corporation||System and method for network interfacing|
|US7936767||Apr 30, 2007||May 3, 2011||International Business Machines Corporation||Systems and methods for monitoring high speed network traffic via sequentially multiplexed data streams|
|US7953103||Nov 24, 2009||May 31, 2011||Cisco Technology, Inc.||Multi-homing using controlled route leakage at a backup service provider|
|US7953903||May 31, 2011||Habanero Holdings, Inc.||Real time detection of changed resources for provisioning and management of fabric-backplane enterprise servers|
|US7974278||Jul 5, 2011||Integrated Device Technology, Inc.||Packet switch with configurable virtual channels|
|US7990994||Oct 22, 2005||Aug 2, 2011||Habanero Holdings, Inc.||Storage gateway provisioning and configuring|
|US7991875 *||Jan 6, 2006||Aug 2, 2011||Intel Corporation||Link level retry scheme|
|US7995696 *||Aug 9, 2011||Integrated Device Technology, Inc.||System and method for deskewing data transmitted through data lanes|
|US8010707 *||Aug 30, 2011||Broadcom Corporation||System and method for network interfacing|
|US8014288||Nov 6, 2008||Sep 6, 2011||Integrated Device Technology, Inc.||Packet latency based arbitration technique for a packet switch|
|US8031609 *||Oct 19, 2009||Oct 4, 2011||Broadcom Corporation||Apparatus and method for communicating arbitrarily encoded data over a 1-gigabit ethernet|
|US8050260 *||Jan 30, 2007||Nov 1, 2011||Qlogic, Corporation||Method and system for load balancing in infiniband switches and networks|
|US8081646||Dec 20, 2011||Integrated Device Technology, Inc.||Old virtual queues technique for routing data packets in a packet switch|
|US8081650 *||Dec 20, 2011||Qlogic, Corporation||Method and system for selecting virtual lanes in fibre channel switches|
|US8089902 *||Jan 3, 2012||Pmc-Sierra Us, Inc.||Serial attached SCSI broadcast primitive processor filtering for loop architectures|
|US8094569||Jan 10, 2012||Cisco Technology, Inc.||Failover and failback of communication between a router and a network switch|
|US8099648 *||Jan 17, 2012||Silicon Image, Inc.||Error detection in physical interfaces for point-to-point communications between integrated circuits|
|US8103664||Jan 24, 2012||Triumfant, Inc.||System for automated computer support|
|US8104087||Jan 24, 2012||Triumfant, Inc.||Systems and methods for automated data anomaly correction in a computer network|
|US8145785||Mar 27, 2012||Habanero Holdings, Inc.||Unused resource recognition in real time for provisioning and management of fabric-backplane enterprise servers|
|US8161185 *||Apr 17, 2012||Cisco Technology, Inc.||Method and apparatus for assigning IPv6 link state identifiers|
|US8161210||Apr 17, 2012||Integrated Device Technology Inc.||Multi-queue system and method for deskewing symbols in data streams|
|US8171121||Sep 23, 2008||May 1, 2012||Intel Corporation||Method, system, and apparatus for dynamic reconfiguration of resources|
|US8190699 *||Jul 28, 2008||May 29, 2012||Crossfield Technology LLC||System and method of multi-path data communications|
|US8190722||May 29, 2012||Randy Oyadomari||Synchronization of timestamps to compensate for communication latency between devices|
|US8190960 *||Dec 5, 2008||May 29, 2012||Force10 Networks, Inc.||Guaranteed inter-process communication|
|US8196013 *||Jun 5, 2012||Plx Technology, Inc.||Supporting global input/output interconnect features on ports of a midpoint device|
|US8219747 *||Jul 10, 2012||Hitachi, Ltd.||Storage system and communications method|
|US8223745 *||Dec 1, 2005||Jul 17, 2012||Oracle America, Inc.||Adding packet routing information without ECRC recalculation|
|US8249846||Aug 21, 2012||International Business Machines Corporation||Automated simulation fabric discovery and configuration|
|US8264953||Sep 11, 2012||Harris Stratex Networks, Inc.||Resilient data communications with physical layer link aggregation, extended failure detection and load balancing|
|US8264959||Jun 15, 2010||Sep 11, 2012||Harris Stratex Networks Operating Corporation||Dynamic load balancing for layer-2 link aggregation|
|US8266271 *||Sep 11, 2012||Jds Uniphase Corporation||Propagation of signals between devices for triggering capture of network data|
|US8271620 *||Nov 13, 2003||Sep 18, 2012||Lantronix, Inc.||Communication protocol converter and method of protocol conversion|
|US8275922 *||Sep 25, 2012||International Business Machines Corporation||Implementing serial link training patterns separated by random data for training a serial link in an interconnect system|
|US8285884||Nov 3, 2008||Oct 9, 2012||Integrated Device Technology Inc.||Data aggregation system and method for deskewing data at selectable data rates|
|US8295299||Oct 1, 2004||Oct 23, 2012||Qlogic, Corporation||High speed fibre channel switch element|
|US8327113||Dec 4, 2012||Intel Corporation||Method, system, and apparatus for dynamic reconfiguration of resources|
|US8335157 *||May 17, 2010||Dec 18, 2012||Cisco Technology, Inc.||Adaptive queue-management|
|US8335756 *||Dec 21, 2007||Dec 18, 2012||Cogniva Information Solutions Inc.||Software for facet classification and information management|
|US8352797 *||Jan 8, 2013||Microsoft Corporation||Software fault isolation using byte-granularity memory protection|
|US8369233 *||Oct 2, 2008||Feb 5, 2013||Endace Technology Limited||Lane synchronisation|
|US8385374 *||Jun 15, 2010||Feb 26, 2013||Marvell Israel (M.I.S.L.) Ltd.||Multilane communication device|
|US8443066||Jul 7, 2011||May 14, 2013||Oracle International Corporation||Programmatic instantiation, and provisioning of servers|
|US8458390||Mar 26, 2012||Jun 4, 2013||Oracle International Corporation||Methods and systems for handling inter-process and inter-module communications in servers and server clusters|
|US8463880 *||Jun 11, 2013||Hewlett-Packard Development Company, L.P.||Method and system for removing a tunnel between portal points|
|US8483108||Jul 24, 2007||Jul 9, 2013||Apple Inc.||Apparatus and methods for de-emphasis training on a point-to-point connection|
|US8539119||Nov 23, 2005||Sep 17, 2013||Qualcomm Incorporated||Methods and apparatus for exchanging messages having a digital data interface device message format|
|US8539149||Jul 10, 2012||Sep 17, 2013||Hitachi, Ltd.||Storage system and communications method|
|US8543890||Apr 26, 2012||Sep 24, 2013||Plx Technology, Inc.||Supporting global input/output interconnect features on ports of a midpoint device|
|US8566473 *||Feb 27, 2009||Oct 22, 2013||Intel Corporation||General input/output architecture, protocol and related methods to implement flow control|
|US8576820||Jul 2, 2004||Nov 5, 2013||Broadcom Corporation||Standby mode for use in a device having a multiple channel physical layer|
|US8578069 *||Apr 4, 2012||Nov 5, 2013||International Business Machines Corporation||Prefetching for a shared direct memory access (DMA) engine|
|US8601053||Jan 17, 2011||Dec 3, 2013||Oracle International Corporation||Multi-chassis fabric-backplane enterprise servers|
|US8605575||Nov 14, 2011||Dec 10, 2013||Cisco Technology, Inc.||Failover and failback of communication between a router and a network switch|
|US8606934 *||Jan 5, 2009||Dec 10, 2013||Intel Corporation||Method, system, and apparatus for system level initialization by conveying capabilities and identifiers of components|
|US8606946||Nov 12, 2004||Dec 10, 2013||Qualcomm Incorporated||Method, system and computer program for driving a data signal in data interface communication data link|
|US8611215||Sep 22, 2011||Dec 17, 2013||Qualcomm Incorporated||Systems and methods for digital data transmission rate control|
|US8625625||Apr 26, 2011||Jan 7, 2014||Qualcomm Incorporated||High data rate interface apparatus and method|
|US8630305||Jun 3, 2005||Jan 14, 2014||Qualcomm Incorporated||High data rate interface apparatus and method|
|US8630318||Jan 29, 2010||Jan 14, 2014||Qualcomm Incorporated||High data rate interface apparatus and method|
|US8630358 *||Mar 20, 2012||Jan 14, 2014||Arm Limited||Data packet flow control across an asynchronous clock domain boundary|
|US8635358||Oct 5, 2010||Jan 21, 2014||Qualcomm Incorporated||High data rate interface|
|US8645566||Mar 24, 2005||Feb 4, 2014||Qualcomm Incorporated||High data rate interface apparatus and method|
|US8645601||May 31, 2011||Feb 4, 2014||Smsc Holdings S.A.R.L.||Methods and systems for performing serial data communication between a host device and a connected device|
|US8650304||Apr 14, 2005||Feb 11, 2014||Qualcomm Incorporated||Determining a pre skew and post skew calibration data rate in a mobile display digital interface (MDDI) communication system|
|US8665723 *||Aug 30, 2011||Mar 4, 2014||International Business Machines Corporation||Managing data on Infiniband (IB) networks|
|US8667363||Nov 23, 2005||Mar 4, 2014||Qualcomm Incorporated||Systems and methods for implementing cyclic redundancy checks|
|US8669988||Mar 10, 2005||Mar 11, 2014||Qualcomm Incorporated||High data rate interface apparatus and method|
|US8670457||Dec 8, 2004||Mar 11, 2014||Qualcomm Incorporated||High data rate interface with improved link synchronization|
|US8681817||Oct 28, 2008||Mar 25, 2014||Qualcomm Incorporated||Generating and implementing a signal protocol and interface for higher data rates|
|US8687658||Nov 24, 2004||Apr 1, 2014||Qualcomm Incorporated||High data rate interface with improved link synchronization|
|US8692838||Nov 23, 2005||Apr 8, 2014||Qualcomm Incorporated||Methods and systems for updating a buffer|
|US8692839||Jan 25, 2008||Apr 8, 2014||Qualcomm Incorporated||Methods and systems for updating a buffer|
|US8693308 *||Feb 10, 2006||Apr 8, 2014||Aviat U.S., Inc.||System and method for resilient wireless packet communications|
|US8694652||Oct 15, 2004||Apr 8, 2014||Qualcomm Incorporated||Method, system and computer program for adding a field to a client capability packet sent from a client to a host|
|US8694663||Jul 15, 2010||Apr 8, 2014||Qualcomm Incorporated||System for transferring digital data at a high rate between a host and a client over a communication path for presentation to a user|
|US8699330||Nov 23, 2005||Apr 15, 2014||Qualcomm Incorporated||Systems and methods for digital data transmission rate control|
|US8700744||Oct 28, 2008||Apr 15, 2014||Qualcomm Incorporated||Generating and implementing a signal protocol and interface for higher data rates|
|US8705521||Mar 17, 2005||Apr 22, 2014||Qualcomm Incorporated||High data rate interface apparatus and method|
|US8705571||Aug 12, 2004||Apr 22, 2014||Qualcomm Incorporated||Signal interface for higher data rates|
|US8705579||Jun 2, 2004||Apr 22, 2014||Qualcomm Incorporated||Generating and implementing a signal protocol and interface for higher data rates|
|US8713295||Apr 17, 2007||Apr 29, 2014||Oracle International Corporation||Fabric-backplane enterprise servers with pluggable I/O sub-system|
|US8719334 *||Sep 10, 2004||May 6, 2014||Qualcomm Incorporated||High data rate interface|
|US8723705||Aug 8, 2006||May 13, 2014||Qualcomm Incorporated||Low output skew double data rate serial encoder|
|US8730069||Nov 9, 2007||May 20, 2014||Qualcomm Incorporated||Double data rate serial encoder|
|US8730913||Apr 26, 2011||May 20, 2014||Qualcomm Incorporated||High data rate interface apparatus and method|
|US8730991 *||Jan 28, 2009||May 20, 2014||Smsc Holdings S.A.R.L.||Device identification using frequency modulated signal bursts|
|US8743872||Jul 9, 2012||Jun 3, 2014||Oracle International Corporation||Storage traffic communication via a switch fabric in accordance with a VLAN|
|US8743878 *||Aug 30, 2011||Jun 3, 2014||International Business Machines Corporation||Path resolve in symmetric infiniband networks|
|US8745251||Apr 16, 2004||Jun 3, 2014||Qualcomm Incorporated||Power reduction system for an apparatus for high data rate signal transfer using a communication protocol|
|US8756294||Oct 29, 2004||Jun 17, 2014||Qualcomm Incorporated||High data rate interface|
|US8774000||Nov 28, 2011||Jul 8, 2014||Harris Stratex Networks, Inc.||Resilient data communications with physical layer link aggregation, extended failure detection and load balancing|
|US8782302 *||Dec 15, 2011||Jul 15, 2014||Stmicroelectronics (Grenoble 2) Sas||Method and apparatus for routing transactions through partitions of a system-on-chip|
|US8804960||Feb 22, 2010||Aug 12, 2014||International Business Machines Corporation||Implementing known scrambling relationship among multiple serial links|
|US8812706||Sep 6, 2002||Aug 19, 2014||Qualcomm Incorporated||Method and apparatus for compensating for mismatched delays in signals of a mobile display interface (MDDI) system|
|US8819005||Jan 3, 2012||Aug 26, 2014||Triumfant, Inc.||System for automated computer support|
|US8819306||Dec 28, 2012||Aug 26, 2014||Intel Corporation||General input/output architecture with PCI express protocol with credit-based flow control|
|US8832216 *||Aug 31, 2011||Sep 9, 2014||Oracle International Corporation||Method and system for conditional remote direct memory access write|
|US8848727||Dec 11, 2009||Sep 30, 2014||Oracle International Corporation||Hierarchical transport protocol stack for data transfer between enterprise servers|
|US8868790||Feb 12, 2005||Oct 21, 2014||Oracle International Corporation||Processor-memory module performance acceleration in fabric-backplane enterprise servers|
|US8873584||Nov 23, 2005||Oct 28, 2014||Qualcomm Incorporated||Digital data interface device|
|US8898554 *||Dec 28, 2012||Nov 25, 2014||Seoul National Unviersity of Technology Center for Industry Collaboration||NoC-based adaptive error correction apparatus|
|US8917601||Feb 27, 2013||Dec 23, 2014||International Business Machines Corporation||Managing data on InfiniBand (IB) networks|
|US8918307||Mar 12, 2009||Dec 23, 2014||International Business Machines Corporation||Automated system latency detection for fabric simulation|
|US8923120 *||Dec 6, 2012||Dec 30, 2014||Cisco Technology, Inc.||Adaptive queue-management|
|US8964547||Jun 8, 2004||Feb 24, 2015||Oracle America, Inc.||Credit announcement|
|US8964754||Nov 8, 2013||Feb 24, 2015||Foundry Networks, Llc||Backplane interface adapter with error control and redundant fabric|
|US8973074 *||May 9, 2011||Mar 3, 2015||Samsung Electronics Co., Ltd.||Method and system for isochronous communication in audio/video networks|
|US8988981||Feb 18, 2014||Mar 24, 2015||Aviat U.S., Inc.||System and method for resilient wireless packet communications|
|US9003466||Apr 20, 2011||Apr 7, 2015||Samsung Electronics Co., Ltd.||Method and system for isochronous data stream management in high speed audio/video networks|
|US9025475 *||Jan 16, 2012||May 5, 2015||Amazon Technologies, Inc.||Proactively retransmitting data packets in a low latency packet data network|
|US9030937 *||Jul 11, 2013||May 12, 2015||Foundry Networks, Llc||Backplane interface adapter with error control and redundant fabric|
|US9030943||Jul 12, 2012||May 12, 2015||Foundry Networks, Llc||Recovering from failures without impact on data traffic in a shared bus architecture|
|US9042363||Sep 23, 2013||May 26, 2015||Broadcom Corporation||Standby mode for use in a device having a multiple channel physical layer|
|US9049125||Dec 28, 2012||Jun 2, 2015||Intel Corporation||General input/output architecture, protocol and related methods to implement flow control|
|US9071528 *||Dec 28, 2012||Jun 30, 2015||Intel Corporation||General input/output architecture, protocol and related methods to implement flow control|
|US9088495 *||Dec 28, 2012||Jul 21, 2015||Intel Corporation||General input/output architecture, protocol and related methods to implement flow control|
|US9088569||May 12, 2011||Jul 21, 2015||International Business Machines Corporation||Managing access to a shared resource using client access credentials|
|US9110895 *||Jun 30, 2010||Aug 18, 2015||Hewlett-Packard Development Company, L.P.||System and method for a serialized data service|
|US9112780||Feb 13, 2013||Aug 18, 2015||Foundry Networks, Llc||Techniques for processing incoming failure detection protocol packets|
|US9166818||Nov 18, 2013||Oct 20, 2015||Brocade Communications Systems, Inc.||Provisioning single or multistage networks using ethernet service instances (ESIs)|
|US9183041 *||Sep 21, 2012||Nov 10, 2015||International Business Machines Corporation||Input/output traffic backpressure prediction|
|US9183042 *||Nov 11, 2013||Nov 10, 2015||International Business Machines Corporation||Input/output traffic backpressure prediction|
|US9208181 *||Oct 24, 2014||Dec 8, 2015||Netapp, Inc.||Migrating data from legacy storage systems to object storage systems|
|US9223738||Nov 2, 2012||Dec 29, 2015||Intel Corporation||Method, system, and apparatus for dynamic reconfiguration of resources|
|US9280507||Mar 27, 2013||Mar 8, 2016||Intel Corporation||High performance interconnect physical layer|
|US9294943||May 19, 2014||Mar 22, 2016||Harris Stratex Networks, Inc.||Resilient data communications with physical layer link aggregation, extended failure detection and load balancing|
|US9306916 *||Dec 25, 2013||Apr 5, 2016||Cavium, Inc.||System and a method for a remote direct memory access over converged ethernet|
|US9307053 *||Jun 19, 2013||Apr 5, 2016||Cisco Technology, Inc.||Direct data placement over user datagram protocol in a network environment|
|US9311268 *||Oct 21, 2013||Apr 12, 2016||Qlogic, Corporation||Method and system for communication with peripheral devices|
|US9313029 *||Dec 15, 2013||Apr 12, 2016||Cavium, Inc.||Virtualized network interface for remote direct memory access over converged ethernet|
|US9319207 *||Nov 26, 2012||Apr 19, 2016||At&T Intellectual Property I, L.P.||System and method for windowing in full-duplex communications|
|US9331922 *||Aug 10, 2012||May 3, 2016||Broadcom Corporation||Automatic recover after loss of signal event in a network device|
|US9338100||Jun 24, 2013||May 10, 2016||Foundry Networks, Llc||Method and apparatus for aggregating input data streams|
|US9354984||Aug 25, 2014||May 31, 2016||Triumfant, Inc.||System for automated computer support|
|US9361405||Jun 30, 2010||Jun 7, 2016||Hewlett Packard Enterprise Development Lp||System and method for service recommendation service|
|US9374305 *||Oct 24, 2013||Jun 21, 2016||Dell Products L.P.||Packet transfer system|
|US9378005||Apr 12, 2013||Jun 28, 2016||Foundry Networks, Llc||Hitless software upgrades|
|US9389895 *||Dec 17, 2009||Jul 12, 2016||Microsoft Technology Licensing, Llc||Virtual storage target offload techniques|
|US20020059451 *||Aug 23, 2001||May 16, 2002||Yaron Haviv||System and method for highly scalable high-speed content-based filtering and load balancing in interconnected fabrics|
|US20020150099 *||Apr 12, 2002||Oct 17, 2002||Pung Hung Keng||Multicast routing method satisfying quality of service constraints, software and devices|
|US20030018780 *||Jun 19, 2002||Jan 23, 2003||Masanori Kawashima||Method and apparatus for managing network devices|
|US20030086424 *||Aug 16, 2002||May 8, 2003||Nec Corporation||Data transfer apparatus and data transfer method|
|US20030093255 *||Nov 9, 2001||May 15, 2003||Freyensee James P.||Hot plug and hot pull system simulation|
|US20030105877 *||Nov 13, 2001||Jun 5, 2003||Riko Yagiu||Data distribution control device and data distribution control method|
|US20030115364 *||Dec 19, 2001||Jun 19, 2003||Li Shu||Camouflage of network traffic to resist attack|
|US20030131090 *||Mar 29, 2001||Jul 10, 2003||Noriyuki Komiya||Network management system|
|US20030140099 *||Nov 8, 2002||Jul 24, 2003||International Business Machines Corporation||Method and apparatus for hard address conflict resolution for enclosures in a loop network|
|US20030154112 *||Oct 29, 2002||Aug 14, 2003||Steven Neiman||System and method for allocating computing resources|
|US20030156594 *||Sep 30, 2002||Aug 21, 2003||Trott Mitchell D.||Slot structure for radio communications system|
|US20030158992 *||Aug 23, 2002||Aug 21, 2003||Jasmin Ajanovic||General input/output architecture, protocol and related methods to implement flow control|
|US20030193894 *||Apr 12, 2002||Oct 16, 2003||Tucker S. Paul||Method and apparatus for early zero-credit determination in an infiniband system|
|US20030231642 *||Apr 2, 2003||Dec 18, 2003||Guiquan Mao||Data upgrade method for a switching device in two-layer network environment|
|US20040015968 *||Jun 20, 2002||Jan 22, 2004||Steven Neiman||System architecture for distributed computing and method of using the system|
|US20040062242 *||Sep 26, 2002||Apr 1, 2004||Wadia Percy K.||Routing packets in packet-based input/output communications|
|US20040093411 *||Aug 29, 2003||May 13, 2004||Uri Elzur||System and method for network interfacing|
|US20040190545 *||Mar 29, 2003||Sep 30, 2004||Reid Geoffrey A.||Soft addressing for disk drive enclosures|
|US20040199652 *||Apr 16, 2004||Oct 7, 2004||Qiuzhen Zou||Generating and implementing a communication protocol and interface for high data rate signal transfer|
|US20040220951 *||Jun 1, 2004||Nov 4, 2004||Howard John H.||System and method for synchronizing mirrored and striped disk writes|
|US20040225720 *||May 9, 2003||Nov 11, 2004||Pinkerton James T.||Remote invalidation of pre-shared RDMA key|
|US20040236624 *||May 22, 2003||Nov 25, 2004||International Business Machines Corporation||Method and apparatus for targeted marketing in embedded chip post issuance transactions|
|US20040247313 *||May 23, 2002||Dec 9, 2004||Masao Tanaka||Media converter|
|US20050010691 *||Jun 30, 2004||Jan 13, 2005||Randy Oyadomari||Synchronization of timestamps to compensate for communication latency between devices|
|US20050013258 *||Jul 12, 2004||Jan 20, 2005||Fike John M.||Method and apparatus for detecting and removing orphaned primitives in a fibre channel network|
|US20050013318 *||Jul 12, 2004||Jan 20, 2005||Fike John M.||Method and system for fibre channel arbitrated loop acceleration|
|US20050013609 *||Jul 12, 2004||Jan 20, 2005||Fike John M.||Method and system for minimizing disruption in common-access networks|
|US20050015518 *||Jul 12, 2004||Jan 20, 2005||Wen William J.||Method and system for non-disruptive data capture in networks|
|US20050018603 *||Jul 20, 2004||Jan 27, 2005||Dropps Frank R.||Method and system for reducing latency and congestion in fibre channel switches|
|US20050018604 *||Jul 20, 2004||Jan 27, 2005||Dropps Frank R.||Method and system for congestion control in a fibre channel switch|
|US20050018606 *||Jul 20, 2004||Jan 27, 2005||Dropps Frank R.||Method and system for congestion control based on optimum bandwidth allocation in a fibre channel switch|
|US20050018621 *||Jul 20, 2004||Jan 27, 2005||Dropps Frank R.||Method and system for selecting virtual lanes in fibre channel switches|
|US20050018649 *||Jul 20, 2004||Jan 27, 2005||Dropps Frank R.||Method and system for improving bandwidth and reducing idles in fibre channel switches|
|US20050018650 *||Jul 20, 2004||Jan 27, 2005||Dropps Frank R.||Method and system for configuring fibre channel ports|
|US20050018663 *||Jul 20, 2004||Jan 27, 2005||Dropps Frank R.||Method and system for power control of fibre channel switches|
|US20050018672 *||Jul 20, 2004||Jan 27, 2005||Dropps Frank R.||Lun based hard zoning in fibre channel switches|
|US20050018674 *||Jul 20, 2004||Jan 27, 2005||Dropps Frank R.||Method and system for buffer-to-buffer credit recovery in fibre channel systems using virtual and/or pseudo virtual lanes|
|US20050018675 *||Jul 20, 2004||Jan 27, 2005||Dropps Frank R.||Multi-speed cut through operation in fibre channel|
|US20050018676 *||Jul 20, 2004||Jan 27, 2005||Dropps Frank R.||Programmable pseudo virtual lanes for fibre channel systems|
|US20050018680 *||Jul 20, 2004||Jan 27, 2005||Dropps Frank R.||Method and system for programmable data dependant network routing|
|US20050018701 *||Jul 20, 2004||Jan 27, 2005||Dropps Frank R.||Method and system for routing fibre channel frames|
|US20050021885 *||Jun 2, 2004||Jan 27, 2005||Anderson Jon James||Generating and implementing a signal protocol and interface for higher data rates|
|US20050022100 *||Jul 22, 2003||Jan 27, 2005||Intel Corporation||Retraining derived clock receivers|
|US20050025060 *||Jul 12, 2004||Feb 3, 2005||Fike John M.||Method and apparatus for testing loop pathway integrity in a fibre channel arbitrated loop|
|US20050025193 *||Jul 12, 2004||Feb 3, 2005||Fike John M.||Method and apparatus for test pattern generation|
|US20050027877 *||Jul 12, 2004||Feb 3, 2005||Fike Melanie A.||Method and apparatus for accelerating receive-modify-send frames in a fibre channel network|
|US20050030808 *||Jul 2, 2004||Feb 10, 2005||Broadcom Corporation||Standby mode for use in a device having a multiple channel physical layer|
|US20050030893 *||Jul 20, 2004||Feb 10, 2005||Dropps Frank R.||Method and system for detecting congestion and over subscription in a fibre channel network|
|US20050030954 *||Jul 20, 2004||Feb 10, 2005||Dropps Frank R.||Method and system for programmable data dependant network routing|
|US20050030978 *||Jul 20, 2004||Feb 10, 2005||Dropps Frank R.||Method and system for managing traffic in fibre channel systems|
|US20050038949 *||Jan 6, 2004||Feb 17, 2005||Lynn Patterson||Apparatus for enabling distributed processing across a plurality of circuit cards|
|US20050044267 *||Jul 20, 2004||Feb 24, 2005||Dropps Frank R.||Method and system for routing and filtering network data packets in fibre channel systems|
|US20050060402 *||Jun 30, 2004||Mar 17, 2005||Randy Oyadomari||Propagation of signals between devices for triggering capture of network data|
|US20050060413 *||Jun 14, 2004||Mar 17, 2005||Randy Oyadomari||Discovery and self-organization of topology in multi-chassis systems|
|US20050080903 *||Sep 30, 2003||Apr 14, 2005||Moshe Valenci||Method, system, and program for maintaining a link between two network entities|
|US20050086349 *||Oct 16, 2003||Apr 21, 2005||Nagarajan Subramaniyan||Methods and apparatus for offloading TCP/IP processing using a protocol driver interface filter driver|
|US20050106941 *||Nov 13, 2003||May 19, 2005||Witchey Nicholas J.||Communication protocol converter and method of protocol conversion|
|US20050117601 *||Aug 12, 2004||Jun 2, 2005||Anderson Jon J.||Signal interface for higher data rates|
|US20050120079 *||Sep 10, 2004||Jun 2, 2005||Anderson Jon J.||High data rate interface|
|US20050125840 *||Oct 15, 2004||Jun 9, 2005||Anderson Jon J.||High data rate interface|
|US20050135251 *||Feb 15, 2005||Jun 23, 2005||Kunz James A.||Method and system for reducing congestion in computer networks|
|US20050135390 *||Nov 12, 2004||Jun 23, 2005||Anderson Jon J.||High data rate interface with improved link control|
|US20050138291 *||Jan 21, 2005||Jun 23, 2005||Steven Neiman||System and method for caching results|
|US20050144225 *||Oct 29, 2004||Jun 30, 2005||Anderson Jon J.||High data rate interface|
|US20050144313 *||Nov 20, 2003||Jun 30, 2005||International Business Machines Corporation||Infiniband multicast operation in an LPAR environment|
|US20050152386 *||Jan 12, 2004||Jul 14, 2005||Lesartre Gregg B.||Successful transactions|
|US20050155040 *||Sep 3, 2004||Jul 14, 2005||Sanjiv Doshi||Implicit interprocess communications (IPC) versioning support|
|US20050163116 *||Nov 24, 2004||Jul 28, 2005||Anderson Jon J.||High data rate interface with improved link synchronization|
|US20050165955 *||Dec 20, 2004||Jul 28, 2005||Duncan Wakelin||Method of using a storage switch and apparatus using and controlling same|
|US20050174936 *||Mar 11, 2004||Aug 11, 2005||Betker Steven M.||Method and system for preventing deadlock in fibre channel fabrics using frame priorities|
|US20050174942 *||Mar 11, 2004||Aug 11, 2005||Betker Steven M.||Method and system for reducing deadlock in fibre channel fabrics using virtual lanes|
|US20050204057 *||Dec 8, 2004||Sep 15, 2005||Anderson Jon J.||High data rate interface with improved link synchronization|
|US20050213585 *||Mar 26, 2004||Sep 29, 2005||Samsung Electronics Co., Ltd.||Apparatus and method for distributing forwarding table lookup operations among a plurality of microengines in a high-speed routing node|
|US20050213593 *||Mar 10, 2005||Sep 29, 2005||Anderson Jon J||High data rate interface apparatus and method|
|US20050216599 *||Mar 17, 2005||Sep 29, 2005||Anderson Jon J||High data rate interface apparatus and method|
|US20050238353 *||Oct 8, 2004||Oct 27, 2005||Mcglaughlin Edward C||Fibre channel transparent switch for mixed switch fabrics|
|US20050259670 *||Mar 24, 2005||Nov 24, 2005||Anderson Jon J||High data rate interface apparatus and method|
|US20050262184 *||May 21, 2004||Nov 24, 2005||Naveen Cherukuri||Method and apparatus for interactively training links in a lockstep fashion|
|US20050262280 *||May 21, 2004||Nov 24, 2005||Naveen Cherukuri||Method and apparatus for acknowledgement-based handshake mechanism for interactively training links|
|US20050268201 *||May 26, 2004||Dec 1, 2005||Jean Khawand||Method and system for tracking sequence numbers|
|US20050271072 *||Jun 3, 2005||Dec 8, 2005||Anderson Jon J||High data rate interface apparatus and method|
|US20050271073 *||Jun 8, 2004||Dec 8, 2005||Johnsen Bjorn D||Switch method and apparatus with cut-through routing for use in a communications network|
|US20050278704 *||Jun 10, 2004||Dec 15, 2005||Xiotech Corporation||Method, apparatus, and program storage device for detecting failures in data flow in high-availability storage systems|
|US20060002385 *||Jun 8, 2004||Jan 5, 2006||Johnsen Bjorn D||Switching method and apparatus for use in a communications network|
|US20060013397 *||Jul 11, 2005||Jan 19, 2006||International Business Machines Corporation||Channel adapter managed trusted queue pairs|
|US20060018330 *||Jun 30, 2004||Jan 26, 2006||Intel Corporation||Method, system, and program for managing memory requests by devices|
|US20060020725 *||Jul 20, 2004||Jan 26, 2006||Dropps Frank R||Integrated fibre channel fabric controller|
|US20060031842 *||Sep 8, 2005||Feb 9, 2006||Steven Neiman||System and method for dividing computations|
|US20060034326 *||Apr 14, 2005||Feb 16, 2006||Anderson Jon J||High data rate interface apparatus and method|
|US20060045031 *||Sep 2, 2004||Mar 2, 2006||International Business Machines Corporation||Automatic hardware data link initialization using multiple state machines|
|US20060047852 *||Oct 1, 2004||Mar 2, 2006||Shishir Shah||Method and system for using boot servers in networks|
|US20060072473 *||Oct 1, 2004||Apr 6, 2006||Dropps Frank R||High speed fibre channel switch element|
|US20060072580 *||Oct 1, 2004||Apr 6, 2006||Dropps Frank R||Method and system for transferring data drectly between storage devices in a storage area network|
|US20060072616 *||Oct 1, 2004||Apr 6, 2006||Dropps Frank R||Method and system for LUN remapping in fibre channel networks|
|US20060092943 *||Nov 4, 2004||May 4, 2006||Cisco Technology, Inc.||Method and apparatus for guaranteed in-order delivery for FICON over SONET/SDH transport|
|US20060133417 *||Dec 2, 2004||Jun 22, 2006||International Business Machines Corporation||Method and system for providing interoperability of different version protocols without access to protocol specific code|
|US20060159023 *||Jan 14, 2005||Jul 20, 2006||International Business Machines Corporation||CRC error history mechanism|
|US20060159081 *||Jan 18, 2005||Jul 20, 2006||Dropps Frank R||Address translation in fibre channel switches|
|US20060168496 *||Nov 23, 2005||Jul 27, 2006||Brian Steele||Systems and methods for implementing cyclic redundancy checks|
|US20060171414 *||Nov 23, 2005||Aug 3, 2006||Behnam Katibian||Systems and methods for digital data transmission rate control|
|US20060179164 *||Nov 23, 2005||Aug 10, 2006||Behnam Katibian||Digital data interface device message format|
|US20060184606 *||Feb 11, 2005||Aug 17, 2006||Lesartre Gregg B||Passing identification information|
|US20060184831 *||Feb 11, 2005||Aug 17, 2006||Lesartre Gregg B||Passing debug information|
|US20060218316 *||Mar 22, 2005||Sep 28, 2006||International Business Machines Corporation||System and method for efficient implementation of a shared receive queue|
|US20060239287 *||Dec 1, 2005||Oct 26, 2006||Johnsen Bjorn D||Adding packet routing information without ECRC recalculation|
|US20060242330 *||Dec 1, 2005||Oct 26, 2006||Ola Torudbakken||Proxy-based device sharing|
|US20060242332 *||Dec 1, 2005||Oct 26, 2006||Johnsen Bjorn D||Distributed I/O bridging functionality|
|US20060242333 *||Dec 1, 2005||Oct 26, 2006||Johnsen Bjorn D||Scalable routing and addressing|
|US20060242352 *||Dec 1, 2005||Oct 26, 2006||Ola Torudbakken||Device sharing|
|US20060259570 *||May 13, 2005||Nov 16, 2006||Microsoft Corporation||Method and system for closing an RDMA connection|
|US20060268681 *||May 31, 2005||Nov 30, 2006||Raza Syed K||Multi-homing using controlled route leakage at a backup service provider|
|US20060288133 *||Nov 23, 2005||Dec 21, 2006||Behnam Katibian||Digital data interface device|
|US20070081527 *||Dec 8, 2006||Apr 12, 2007||Betker Steven M||Method and system for primary blade selection in a multi-module fibre channel switch|
|US20070094345 *||Sep 28, 2001||Apr 26, 2007||Rabbers David L||Method and apparatus for detecting insufficient memory for data extraction processes|
|US20070103338 *||Oct 13, 2005||May 10, 2007||Teo Jeff B K||System and method for measuring and correcting data lane skews|
|US20070130353 *||Jan 6, 2006||Jun 7, 2007||Intel Corporation||Link level retry scheme|
|US20070162651 *||Dec 15, 2006||Jul 12, 2007||Arm Limited||Data transfer control|
|US20070189154 *||Feb 10, 2006||Aug 16, 2007||Stratex Networks, Inc.||System and method for resilient wireless packet communications|
|US20070201457 *||May 3, 2007||Aug 30, 2007||Betker Steven M||Method and system for dynamically assigning domain identification in a multi-module fibre channel switch|
|US20070250640 *||Apr 24, 2006||Oct 25, 2007||Cisco Technology, Inc.||Method and apparatus for assigning Ipv6 link state identifiers|
|US20070263647 *||Mar 9, 2007||Nov 15, 2007||Peter Shorty||Using battery-powered nodes in a mesh network|
|US20080005706 *||Jun 30, 2006||Jan 3, 2008||Debendra Das Sharma||Method and apparatus for improving high availability in a PCI express link through predictive failure analysis|
|US20080016511 *||Sep 26, 2007||Jan 17, 2008||Microsoft Corporation||Method and computer program product for offloading processing tasks from software to hardware|
|US20080034137 *||Jul 24, 2007||Feb 7, 2008||Colin Whitby-Strevens||Apparatus and methods for de-emphasis training on a point-to-point connection|
|US20080034160 *||Jul 3, 2007||Feb 7, 2008||Jp Morgan Chase & Co.||System and method for caching results|
|US20080036631 *||Aug 8, 2006||Feb 14, 2008||Qualcomm Incorporated||Low output skew double data rate serial encoder|
|US20080088492 *||Nov 9, 2007||Apr 17, 2008||Qualcomm Incorporated||Double data rate serial encoder|
|US20080117830 *||Nov 20, 2006||May 22, 2008||International Business Machines Corporation||Detection of Link Status in Blade Systems|
|US20080129749 *||Jan 25, 2008||Jun 5, 2008||Qualcomm Incorporated||Methods and systems for updating a buffer|
|US20080168302 *||Jan 10, 2007||Jul 10, 2008||International Business Machines Corporation||Systems and methods for diagnosing faults in a multiple domain storage system|
|US20080175268 *||Mar 25, 2008||Jul 24, 2008||International Business Machines Corporation||method and system for providing interoperability of different version protocols without access to protocol specific code|
|US20080267183 *||Jul 2, 2008||Oct 30, 2008||International Business Machines Corporation||Infiniband Multicast Operation in an LPAR Environment|
|US20080267192 *||Apr 30, 2007||Oct 30, 2008||International Business Machines Corporation||Systems and methods for monitoring high speed network traffic via sequentially multiplexed data streams|
|US20080270638 *||Apr 30, 2007||Oct 30, 2008||International Business Machines Corporation||Systems and methods for monitoring high speed network traffic via simultaneously multiplexed data streams|
|US20090019267 *||Sep 23, 2008||Jan 15, 2009||Mani Ayyar||Method, System, and Apparatus for Dynamic Reconfiguration of Resources|
|US20090024715 *||Sep 23, 2008||Jan 22, 2009||Mani Ayyar||Method, System, and Apparatus for Dynamic Reconfiguration of Resources|
|US20090055600 *||Sep 23, 2008||Feb 26, 2009||Mani Ayyar||Method, System, and Apparatus for Dynamic Reconfiguration of Resources|
|US20090055709 *||Oct 28, 2008||Feb 26, 2009||Qualcomm Incorporated||Generating and implementing a signal protocol and interface for higher data rates|
|US20090067324 *||Sep 5, 2008||Mar 12, 2009||Harris Stratex Networks Operating Corporation||Resilient Data Communications with Physical Layer Link Aggregation, Extended Failure Detection and Load Balancing|
|US20090070479 *||Oct 28, 2008||Mar 12, 2009||Qualcomm Incorporated||Generating and implementing a signal protocol and interface for higher data rates|
|US20090106448 *||Dec 17, 2008||Apr 23, 2009||Riko Yagiu||Data distribution management device and data distribution management method|
|US20090133104 *||Jan 21, 2009||May 21, 2009||Canon Kabushiki Kaisha||Device Access Based on Centralized Authentication|
|US20090154484 *||Dec 11, 2008||Jun 18, 2009||Zarlink Semiconductor Inc.||Scaleable voip telephone line circuit|
|US20090177913 *||Jan 8, 2008||Jul 9, 2009||Triumfant, Inc.||Systems and Methods for Automated Data Anomaly Correction in a Computer Network|
|US20090193164 *||Feb 27, 2009||Jul 30, 2009||Jasmin Ajanovic||General Input/Output Architecture, Protocol and Related Methods to Implement Flow Control|
|US20090240789 *||Mar 24, 2008||Sep 24, 2009||Sankarlingam Dandabany||Method and system for removing a tunnel between portal points|
|US20090254647 *||Jun 8, 2009||Oct 8, 2009||Uri Elzur||System and method for network interfacing|
|US20090265472 *||Jan 5, 2009||Oct 22, 2009||Mani Ayyar||Method, System, and Apparatus for System Level Initialization|
|US20090316592 *||Dec 24, 2009||Dropps Frank R||Method and system for selecting virtual lanes in fibre channel switches|
|US20100005339 *||Aug 27, 2009||Jan 7, 2010||Triumfant, Inc.||System for Automated Computer Support|
|US20100023595 *||Jul 28, 2008||Jan 28, 2010||Crossfield Technology LLC||System and method of multi-path data communications|
|US20100034213 *||Feb 11, 2010||Broadcom Corporation||Apparatus and Method for Communicating Arbitrarily Encoded Data Over a 1-Gigabit Ethernet|
|US20100036789 *||Dec 21, 2007||Feb 11, 2010||Cogniva Information Solutions Inc.||Software for facet classification and information management|
|US20100074268 *||Nov 24, 2009||Mar 25, 2010||Cisco Technology, Inc.||Multi-homing using controlled route leakage at a backup service provider|
|US20100077002 *||Apr 4, 2007||Mar 25, 2010||Knud Funch||Direct access method to media information|
|US20100085873 *||Apr 8, 2010||Marc Maria Moons||Lane synchronisation|
|US20100091621 *||Nov 28, 2008||Apr 15, 2010||Makio Mizuno||Storage System and Communications Method|
|US20100128626 *||Jan 29, 2010||May 27, 2010||Qualcomm Incorporated||High data rate interface apparatus and method|
|US20100142368 *||Dec 5, 2008||Jun 10, 2010||Naveen Gunukula||Failover and failback of communication between a router and a network switch|
|US20100235156 *||Sep 16, 2010||International Business Machines Corporation||Automated Simulation Fabric Discovery and Configuration|
|US20100235158 *||Sep 16, 2010||International Business Machines Corporation||Automated System Latency Detection for Fabric Simulation|
|US20100246396 *||Sep 30, 2010||Sergio Licardie||Dynamic Load Balancing for Layer-2 Link Aggregation|
|US20100260055 *||Aug 10, 2009||Oct 14, 2010||Qualcomm Incorporated||High data rate interface with improved link synchronization|
|US20110013681 *||Jul 15, 2010||Jan 20, 2011||Qualcomm Incorporated||Generating and implementing a communication protocol and interface for high data rate signal transfer|
|US20110022719 *||Jan 27, 2011||Qualcomm Incorporated||High data rate interface|
|US20110099456 *||Oct 23, 2009||Apr 28, 2011||Plx Technology, Inc.||Supporting global input/output interconnect features on ports of a midpoint device|
|US20110138476 *||Jun 9, 2011||Microsoft Corporation||Software Fault Isolation Using Byte-Granularity Memory Protection|
|US20110145640 *||Jun 16, 2011||Triumfant, Inc.||System for Automated Computer Support|
|US20110154318 *||Dec 17, 2009||Jun 23, 2011||Microsoft Corporation||Virtual storage target offload techniques|
|US20110199931 *||Aug 18, 2011||Qualcomm Incorporated||High data rate interface apparatus and method|
|US20110206141 *||Aug 25, 2011||International Business Machines Corporation||Implementing serial link training patterns separated by random data|
|US20110208954 *||Feb 22, 2010||Aug 25, 2011||International Business Machines Corporation||Implementing known scrambling relationship among multiple serial links|
|US20110209027 *||Aug 25, 2011||Silicon Image, Inc.||Error detection in physical interfaces for point-to-point communications between integrated circuits|
|US20110238858 *||Jan 10, 2011||Sep 29, 2011||Hitachi, Ltd.||Traffic shaping device|
|US20110265139 *||Oct 27, 2011||Harkirat Singh||Method and system for isochronous communication in audio/video networks|
|US20110280125 *||Nov 17, 2011||Cisco Technology, Inc.||Adaptive Queue-Management|
|US20110296065 *||Dec 1, 2011||Dirk Daecke||Control unit for the exchange of data with a peripheral unit, peripheral unit, and method for data exchange|
|US20120076151 *||Oct 31, 2011||Mar 29, 2012||Finn Norman W||Multi-Bridge LAN Aggregation|
|US20120150936 *||Jun 2, 2010||Jun 14, 2012||Nec Corporation||Distribution system|
|US20120159017 *||Dec 15, 2011||Jun 21, 2012||Stmicroelectronics Srl||Method and apparatus for routing transactions through partitions of a system-on-chip|
|US20130051232 *||Aug 30, 2011||Feb 28, 2013||International Business Machines Corporation||Managing data on infiniband (ib) networks|
|US20130051394 *||Aug 30, 2011||Feb 28, 2013||International Business Machines Corporation||Path resolve in symmetric infiniband networks|
|US20130054726 *||Feb 28, 2013||Oracle International Corporation||Method and system for conditional remote direct memory access write|
|US20130080411 *||Jun 30, 2010||Mar 28, 2013||Jerome Rolia||System and method for a serialized data service|
|US20130094358 *||Apr 18, 2013||Lenard John Stephan Jayakumar||Adaptive queue-management|
|US20130227066 *||Mar 23, 2012||Aug 29, 2013||Nec Corporation||Data transfer apparatus and data transfer method|
|US20130254451 *||Dec 28, 2012||Sep 26, 2013||Jasmin Ajanovic||General input/output architecture, protocol and related methods to implement flow control|
|US20130254452 *||Dec 28, 2012||Sep 26, 2013||Jasmin Ajanovic||General input/output architecture, protocol and related methods to implement flow control|
|US20130275804 *||Dec 28, 2012||Oct 17, 2013||Seoul National University Of Technology Center For Industry Collaboration||NoC-Based Adaptive Error Correction Apparatus|
|US20140023086 *||Jul 11, 2013||Jan 23, 2014||Foundry Networks, Llc||Backplane Interface Adapter with Error Control and Redundant Fabric|
|US20140043957 *||Aug 10, 2012||Feb 13, 2014||Broadcom Corporation||Automatic Recover After Loss Of Signal Event In A Network Device|
|US20140089607 *||Nov 11, 2013||Mar 27, 2014||International Business Machines Corporation||Input/output traffic backpressure prediction|
|US20140089621 *||Sep 21, 2012||Mar 27, 2014||International Business Machines Corporation||Input/output traffic backpressure prediction|
|US20140112339 *||Oct 22, 2013||Apr 24, 2014||Robert J. Safranek||High performance interconnect|
|US20140114887 *||Mar 15, 2013||Apr 24, 2014||Venkatraman Iyer||High performance interconnect physical layer|
|US20140146715 *||Nov 26, 2012||May 29, 2014||At&T Intellectual Property I, L.P.||System and method for windowing in full-duplex communications|
|US20140211630 *||Sep 28, 2011||Jul 31, 2014||Vincent E. Cavanna||Managing packet flow in a switch faric|
|US20140317075 *||Nov 14, 2013||Oct 23, 2014||Larry Maurice Deutsch||Method and Apparatus to Search Data and Notify and Update a User|
|US20140317165 *||Jun 19, 2013||Oct 23, 2014||Cisco Technology, Inc.||Direct data placement over user datagram protocol in a network environment|
|US20150046502 *||Oct 24, 2014||Feb 12, 2015||Netapp Inc.||Migrating data from legacy storage systems to object storage systems|
|US20150098466 *||Oct 6, 2013||Apr 9, 2015||Mellanox Technologies Ltd.||Simplified packet routing|
|US20150117448 *||Oct 24, 2013||Apr 30, 2015||Dell Products L.P.||Packet transfer system|
|US20150172055 *||Dec 15, 2013||Jun 18, 2015||Cavium, Inc.||Virtualized network interface for remote direct memory access over converged ethernet|
|US20150178242 *||Dec 25, 2013||Jun 25, 2015||Cavium, Inc.||System and a method for a remote direct memory access over converged ethernet|
|US20150180799 *||Dec 20, 2013||Jun 25, 2015||Thomas D. Lovett||Hierarchical/lossless packet preemption to reduce latency jitter in flow-controlled packet-based networks|
|CN101669345B||Feb 26, 2008||May 15, 2013||Ati科技无限责任公司||Robust control/delineation in serial streams|
|CN102201975A *||Dec 20, 2010||Sep 28, 2011||株式会社日立制作所||Traffic shaping device|
|EP2122961A1 *||Feb 26, 2008||Nov 25, 2009||ATI Technologies ULC||Robust control/delineation in serial streams|
|EP2122961A4 *||Feb 26, 2008||Jan 18, 2012||Ati Technologies Ulc||Robust control/delineation in serial streams|
|WO2014065877A1 *||Mar 15, 2013||May 1, 2014||Venkatraman Iyer||High performance interconnect physical layer|
|WO2014065882A1 *||Mar 27, 2013||May 1, 2014||Venkatraman Iyer||High performance interconnect physical layer|
|WO2014065884A1 *||Mar 28, 2013||May 1, 2014||Jeff Willey||High performance interconnect link layer|
|WO2015099724A1 *||Dec 26, 2013||Jul 2, 2015||Intel Corporation||Pci express enhancements|
|U.S. Classification||709/228, 709/237|
|Cooperative Classification||G06F11/2007, H04L1/18, H04L1/24|
|European Classification||H04L1/18, H04L1/24|
|Feb 16, 2001||AS||Assignment|
Owner name: COMPAQ COMPUTER CORPORATION, TEXAS
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:BUNTON, WILLIAM P.;REEL/FRAME:011536/0274
Effective date: 20010212
|Dec 2, 2003||AS||Assignment|
Owner name: HEWLETT-PACKARD DEVELOPMENT COMPANY, L.P., TEXAS
Free format text: CHANGE OF NAME;ASSIGNOR:COMPAQ INFORMATION TECHNOLOGIES GROUP L.P.;REEL/FRAME:014177/0428
Effective date: 20021001
Owner name: HEWLETT-PACKARD DEVELOPMENT COMPANY, L.P.,TEXAS
Free format text: CHANGE OF NAME;ASSIGNOR:COMPAQ INFORMATION TECHNOLOGIES GROUP L.P.;REEL/FRAME:014177/0428
Effective date: 20021001
|May 19, 2009||CC||Certificate of correction|
|Sep 8, 2009||FPAY||Fee payment|
Year of fee payment: 4
|Aug 10, 2012||AS||Assignment|
Owner name: QNAP SYSTEMS, INC., TAIWAN
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:HEWLETT-PACKARD DEVELOPMENT COMPANY, L.P.;HEWLETT-PACKARD COMPANY;REEL/FRAME:028764/0422
Effective date: 20120518
|Mar 15, 2013||FPAY||Fee payment|
Year of fee payment: 8